Fix/episode cache bytes metric wiring

[huynna12]

Description

Closes the loop on #103. The requested tracking (EpisodeCache byte accounting + tether_episode_cache_bytes_total Prometheus gauge) was already implemented in #106. This PR fixes two correctness gaps left in that implementation:

1. Metric never emits in production. The only real call site, Pi05DecomposedInference.__init__ with cache_level="episode" (src/tether/runtime/pi05_decomposed_server.py:378), constructed EpisodeCache without embodiment/model_id. Since EpisodeCache._emit_bytes_metric() no-ops unless both are set, the gauge silently never emitted once episode mode was enabled. Fixed by reusing cuda_graphs_embodiment/cuda_graphs_model_id, already threaded into the same constructor for CUDA graph metrics.

2. insert() double-counts bytes on a same-key re-insert. Calling insert() twice for the same (episode_id, lang_hash) never backed out the old entry's bytes before adding the new ones - a permanent leak in _total_bytes (and therefore the gauge), a chance of evicting an unrelated LRU entry at capacity, and inflated stats.episode_count ("unique episodes seen"). Not reachable via today's only caller (which always checks lookup() first), but insert() is public API on the exact class [Metrics] Track Episode Cache Memory Usage #103 asked to be byte-accurate, so it's fixed directly rather than left latent.

Note: cache_level="episode" isn't wired to any tether serve CLI flag yet, so this doesn't make the feature live today, it makes the byte-tracking and metric correct for when it becomes reachable.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature
• Breaking change
• Documentation update
• Infrastructure / CI update

How Has This Been Tested?

• pytest tests/ passes locally

• tether doctor sanity check

• Other: manual end-to-end script against the real (non-mocked)
  classes, reproducible with:

  pytest tests/test_episode_cache_bytes.py \
         tests/test_episode_cache_production_wiring.py \
         tests/test_episode_cache_utils.py \
         tests/test_decomposed_server.py \
         tests/test_cuda_graphs_integration.py \
         tests/test_pi05_action_fast_path_integration.py -v
  # -> 51 passed, 1 skipped (CUDA-only test, expected without a GPU)

  Confirmed both new test groups fail against the pre-fix code before passing with it:
  assert None == 'franka' (gap 1), assert 1032 == 516 / episode_count 2-vs-1 / evictions 1-vs-0 (gap 2).

  Also drove the real Pi05DecomposedInference class (mocked ORT session only - no real model export available here) through cache_level="episode" and read the actual rendered Prometheus output:

  tether_episode_cache_bytes_total{embodiment="franka",model_id="pi05-real-check",policy_slot="prod"} 516.0
  

• ruff check clean on all touched files.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally with my changes
• I have kept the PR scoped to a single concern