CUDA: promote opt-in exact bitmap join plans

[StarGazerM]

Scope

Promote the reusable opt-in exact bitmap join implementation, not the DOOP research archive or generated query variants.

• Expose with_plan(dedup_bitmap=True) through DSL, HIR/MIR, index planning, and full CUDA runner emission.
• Maintain dictionary indexes across producer strata and recursive iterations; support either recursive source role and physical column order.
• Preserve exact set projection semantics. Reject filters, negation, constants, provenance, count-only/conflicting strategies, unsupported shapes/indexes and unmatched recursive variants instead of approximating them.
• Require complete runner emission; standalone kernel compilation cannot provide dictionary lifecycle management.
• Leave examples/doop.py and default planning unchanged. No automatic optimizer selection or end-to-end speedup claim.

Verification

• Complete integration Python suite: 1,110 passed, 5 skipped.
• Real bounded CUDA smoke: exact exported tuples for 10 relations / 37 tuples, including existing NEW seeds, reversed physical column order, and both recursive source roles; normal process exit. Initial/max RMM pools were 64/256 MiB to avoid disturbing the running vLLM workload.
• Additional native boundary smoke: baseline and bitmap each matched the expected 77,874 tuples exactly, across multiple bitmap words, negative int32 values, overlapping join-key projections and the 65,536-row export-chunk boundary; normal process exit with a 256 MiB pool cap.
• The complete canonical DOOP query compiled successfully in both baseline and bitmap modes with clang++-20/CUDA 12.9.
• Added rejection regressions for specializations that would discard rule semantics and public compilation API coverage.

Full twelve-dataset GPU timing/correctness is not claimed: an existing vLLM process occupies most of the 48 GB GPU. The next stacked PR supplies CPU/GPU execution and exact cross-backend comparison.

Stack

Depends on #91 for review ordering; merge after that PR. This commit contains only compiler/runtime bitmap support and its documentation/tests.