perf(rocm): add host-gap accounting per port unit to the decode profile

[inureyes]

Summary

The decode profile ranked ports by their share of decode GPU time, whose ceiling 1 / (1 - share) ignores the host launch gaps between dispatches. That underrated dispatch-heavy ports: #2067 measured 1.46x and 1.45x against GPU-share ceilings of 1.42x and 1.25x. The profile now charges each idle stretch of the decode window to the dispatch that ends it and gives every port unit a wall-time share and ceiling that count its launch gaps.

• scripts/rocm_decode_gaps.py (new, split out because rocm_decode_profile.py was already over 500 lines): attribute_gaps, the ceilings, and the scaling of traced gaps to the plain run.
• summarize() adds role_host_gap_ms_per_token and, per unit for the fallback and reached-default roles, host gap per token and per dispatch, wall share, ceiling_gpu_share, ceiling_wall and plain_ceiling_wall_est. A host_gap_attribution_residual_ns check (expected 0) pins the per-role gaps to the window's idle time. Existing fields are unchanged; report() adds a ceiling table and shows - for summaries written before this change.
• Docs: docs/benchmarks.md describes the fields and names plain_ceiling_wall_est as the bound to compare a measured speedup against; the 2026-09-30 profile page gets a dated note with the new ceilings.
• The parser still reads the post-epic: one batch-native engine for CLI and server decode #2166 mlxcel-bench-decode output and phase marks (both profiled runs cut cleanly: 0 kernels straddle the decode start), so no parser fix was needed.

Validation of the method (gfx1151, 6dbfe7d7, MLXCEL_SSM_KERNEL=0)

Report output (python3 scripts/rocm_decode_profile.py report benchmarks/rocm_profiles/gfx1151_6dbfe7d7_ssm-graph), ceiling table:

Run: ceiling wall (GPU share)	#2063	#2064	#2065	#2067	#2068
NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_greedy (plain)	1.00 (1.00)	1.00 (1.00)	1.26 (1.37)	1.56 (1.26)	1.00 (1.00)
granite-4.0-h-tiny-4bit_greedy (plain)	1.00 (1.00)	1.00 (1.00)	1.25 (1.29)	1.60 (1.45)	1.00 (1.00)

Measured #2067 speedup on the same binary, median of three alternating runs per arm in one guard session per model: granite 59.88 to 89.40 tok/s, 1.49x; Nemotron-H 52.29 to 75.26 tok/s, 1.44x. Spread under 2% except one granite kernel-off run at 52.18 (median unaffected). Both plain_ceiling_wall_est values (1.60x, 1.56x) are at or above the measured gains, and both GPU-share ceilings (1.45x, 1.26x) are below them. Today's measured gains differ slightly from PR #2099's 1.46x and 1.45x because main has moved since; the comparison uses gains and ceilings from the same commit. Raw results, ab_decode.txt and both guard logs (every attempt CLEAN) are in benchmarks/rocm_profiles/gfx1151_6dbfe7d7_ssm-graph/.

Verification

• On gfx1151 after rebasing onto a18d3d76: make verify-rocm OK, 163 cargo test suites, 12194 passed, 0 failed, 403 ignored, plus 100 Python tooling tests (including the new ones); cargo test --features rocm --test dead_doc_pointers passed.
• make verify-versions verify-kernel-dtype-keys verify-kernel-port-dispatch verify-llama-compat verify-fmt verify-python-tooling passed.

Not verified: Metal and CUDA (not available on this host). The change touches only the ROCm profiling scripts, their tests and docs; no Rust or backend code.

Closes #2148