Caltable: type each site's gains and D-terms from its feed type

[Lftw]

What and why

Every site's gain and D-term tables were typed circular, whatever the station's feeds, so a linear station's X/Y gains were stored under rscale/lscale. Caltable now types each site from tarr['feed_type']:

ct.gains['ALMA']    # DTCAL_LIN:  (time, xscale, yscale)   for an 'xy' station
ct.gains['APEX']    # DTCAL_CIRC: (time, rscale, lscale)   for an 'rl' station
ct.gains[site]['p1scale']   # generic names work on every site

Internally the module reads and writes the generic p1scale/p2scale and d_p1/d_p2 names, so applycal, merge, pad_scans, scan_avg, enforce_positive and invert_gains are one code path for both bases. DTCAL_LIN/DTDTERM_LIN now survive save/load; the file format is unchanged, the basis rides on array.txt. make_jones types its simulated cal table per station instead of always circular.

Pure refactor, no new capability: this is the step before applycal learns to apply gains on any basis (next PR). Circular data is byte-identical throughout.

Design decisions worth discussing

• ehc.caltable_dtypes(feed_type) is the single feed_type -> (DTCAL_*, DTDTERM_*) map. 'rl' and 'xy' only. Hybrid feeds raise like field_rotation_matrix does; '??' says to declare the feeds; 'lr'/'yx' now raise at construction (before, they were silently typed circular, which labelled p1 as R).
• Columns map by name, not position. A (time, lscale, rscale) table used to come out with its feeds swapped. Tables with names that can't be placed in a slot raise instead of being stored and failing later.
• The declared feed type wins over the field names. A circular-typed table on an 'xy' station is relabelled silently, because every solver still writes ehc.DTCAL whatever the basis, so r/l names carry no information. The one direction that warns is x/y names under an 'rl' station, since only basis-aware code writes those.
• plot_gains/plot_compare_gains pol accepts R|L|X|Y|p1|p2|both, case-insensitive. A physical name from the wrong basis raises up front. The default is 'p1', which is 'R' on a circular site.
• merge refuses the same site in two bases; disjoint site lists in different bases (ALMA in LIN, the rest in CIRC) merge as before.

Bugs found along the way

• plot_gains(pol='both') passed the validator and then hit an unset local; it has never worked
• plot_gains bound tmins = tmaxes = gmins = gmaxes = [] to one list, so the y-range was computed from the times as well as the gains
• plot_compare_gains with any pol other than 'R'/'L' hit an unset local
• load_caltable's fallback filename was missing a path separator
• summary_plots.imgsum, modeling_utils.caltable_to_gains and the interactive dashboard read rscale/lscale directly, which a linear station no longer has; they now use the generic names

Deferred

• applycal still forces switch_polrep('circ'); gains on any basis is the next PR, apply_dterms=True the one after
• The solvers (self_cal, network_cal, polgains_cal, modeling_utils) still write ehc.DTCAL; the constructor relabels, so nothing breaks, but they should write the site's own dtype eventually
• make_jones keeps its gainR/gainL variable names: the surrounding gain model is circular-specific (gain_RLratio, rlgaincal), so only the carrier variables would have been renamed

How to test

micromamba run -n jax-ehtim pytest tests/test_caltable.py tests/test_mixedpol.py tests/test_obs_simulate.py

142 + 231 + 95 pass. Sections 1-11 of test_caltable.py are the no-regression contract from #331; the one change there is test_pad_scans_keeps_linear_gain_dtype, whose station is now genuinely 'xy' rather than a LIN table on an 'rl' site, which dispatch makes self-contradictory. Section 12 is new. test_mixedpol.py gains the make_jones cal-table checks; they were confirmed to fail with the writer fix reverted.

Part of #212.

[codecov]

Codecov Report

❌ Patch coverage is 87.73585% with 13 lines in your changes missing coverage. Please review.
✅ Project coverage is 49.40%. Comparing base (571e00f) to head (ea7d82e).

Files with missing lines	Patch %	Lines
ehtim/caltable.py	89.00%	7 Missing and 4 partials ⚠️
ehtim/plotting/interactive.py	66.66%	1 Missing ⚠️
ehtim/plotting/summary_plots.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@                  Coverage Diff                   @@
##           dev-backend-mixpol     #335      +/-   ##
======================================================
+ Coverage               49.17%   49.40%   +0.22%     
======================================================
  Files                      54       54              
  Lines                   27241    27266      +25     
  Branches                 4662     4668       +6     
======================================================
+ Hits                    13397    13471      +74     
+ Misses                  12324    12262      -62     
- Partials                 1520     1533      +13     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.