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

[inureyes]

Part of #1801

Problem / Background

scripts/rocm_decode_profile.py (#2061, PR #2086) ranks the #1814 ports by their share of decode GPU kernel time. That share implies a speedup ceiling of 1/(1 - share), but it ignores the host time between dispatches. For ports that replace many small dispatches the ceiling is wrong: the SSM port (#2067, PR #2099) measured 1.46x decode on granite-4.0-h-tiny and 1.45x on Nemotron-3-Nano, above the 1/(1 - 0.298) = 1.42x and 1/(1 - 0.200) = 1.25x the profile's 29.82% and 20.01% shares allow (docs/benchmark_results/rocm-decode-profile-gfx1151-2026-09-30.md:114,117). The same page reports 3197 and 2008 dispatches per token and 4.8 and 5.1 ms per token of host gap for the two hybrids (line 73), but only as a whole-run number. The next ranking will underrate dispatch-heavy ports the same way.

Current Behavior

• summarize() (scripts/rocm_decode_profile.py:400-521) computes host_gap_ms_per_token, host_gap_pct_of_wall and plain_host_gap_ms_per_token_est for the whole decode window only (lines 489-499).
• Per port unit (UNIT_ROLES, line 350) it records fallback_share_pct, reached_default_share_pct, reached_optin_share_pct and dispatches per token (lines 453-466), all relative to gpu_sum (the sum of kernel durations, share() at line 437).
• report() (lines 544-567) prints the unit table as reached (fallback) GPU-time shares.

Proposed Solution

Attribute idle time to the dispatch that ends it, then derive a wall-time share and ceiling per role and per port unit. Keep the existing GPU-time fields unchanged.

• New pure function attribute_gaps(decode: list[Dispatch], roles: list[str | None], lo: int, hi: int) -> dict[str, int]: walk dispatches in start order, keep frontier = max(lo, max end so far); for each dispatch gap = max(0, start - frontier), added to that dispatch's role ("unattributed" for None). The idle tail from the last end to hi goes to "tail". The total equals wall - gpu_busy from busy_ns().
• Summary additions: top-level role_host_gap_ms_per_token; per unit fallback_host_gap_ms_per_token, fallback_host_gap_us_per_dispatch, fallback_wall_share_pct = 100 * (unit kernel ns + unit gap ns) / wall, ceiling_gpu_share = 1 / (1 - fallback_share_pct / 100) and ceiling_wall = 1 / (1 - fallback_wall_share_pct / 100), both rounded to 2 decimals. When a plain run exists, plain_ceiling_wall_est scales the unit's gap by plain_host_gap_ms_per_token_est / host_gap_ms_per_token (the tracer inflates host gaps, not kernel time) and divides by the plain wall per token. Same fields for the reached-default roles.
• report(): a second unit table with ceiling_wall (plain_ceiling_wall_est when present) and ceiling_gpu_share per run and unit.
• Edge cases: zero tokens leave the per-token fields None as today; a unit with no dispatches reports 0 gap and a ceiling of 1.0; a share of 100% must not divide by zero (report None).
• Docs: in docs/benchmarks.md (section at line 220) say the GPU share ignores launch gaps and that ceiling_wall is the bound to compare a port's measured speedup against. Add a dated note to the 2026-09-30 profile page that the perf(rocm): port ssm_update_kernel to HIP and fix ssm_kernel_available #2067 shares understated the gain, with the new ceilings.

Acceptance Criteria

• Unit tests in tests/test_rocm_decode_profile.py (WindowTests): attribute_gaps on a synthetic trace with overlapping dispatches, a leading gap and an idle tail gives the expected per-role totals, and their sum equals wall - busy_ns(...); summarize() on the existing synthetic trace emits the new keys.
• On gfx1151, scripts/rocm_decode_profile.sh for granite-4.0-h-tiny-4bit and NVIDIA-Nemotron-3-Nano-30B-A3B-4bit with MLXCEL_SSM_KERNEL=0 (the graph path) gives a perf(rocm): port ssm_update_kernel to HIP and fix ssm_kernel_available #2067 plain_ceiling_wall_est at or above the measured 1.46x and 1.45x. If it does not, the PR explains the remaining difference before merging.
• The report output and both docs changes are in the PR.

Verification

python3 -m unittest tests/test_rocm_decode_profile.py -v
cargo build --release --features rocm --bin mlxcel-bench-decode
MLXCEL_SSM_KERNEL=0 scripts/rocm_decode_profile.sh models/mlx/granite-4.0-h-tiny-4bit models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit
python3 scripts/rocm_decode_profile.py report benchmarks/rocm_profiles/gfx1151_<commit>

Related: #2061, PR #2086, #2067, PR #2099, #1814.