perf: scale throughput by effective batch size, not requested --batch-size

[xieofxie]

Summary

Fixes incorrect samples_per_sec in wmk perf when the requested --batch-size cannot be applied to the model.

The requested --batch-size only lands on inputs whose leading dimension is dynamic (see _resolve_shape). For a model with a statically-fixed batch dim (e.g. [1, 3, 224, 224]), input generation silently keeps the static value, so the session runs batch=1 — but throughput was still computed as config.batch_size / latency, inflating samples_per_sec by the requested batch factor (e.g. reporting 800 sps when the model actually processed 100). Latency stats were always correct; only the derived per-sample throughput was wrong.

Changes

• New helper effective_batch_size() — reads the actual batch back from the first batched (rank ≥ 1) generated input, matching the module's existing "first dim is batch" convention. Falls back to the requested value when all inputs are scalar.
• _collect_results() now scales samples_per_sec by the effective batch the session ran, not the requested config.batch_size. (batches_per_sec was already correct — it is per-call.)
• _generate_inputs() logs a warning when the requested --batch-size could not be applied (static batch dim).
• BenchmarkResult gains an effective_batch_size field, surfaced in the JSON benchmark_info alongside the requested batch_size, and shown in the console throughput line with a Note: when the two differ.

Tests

New TestEffectiveBatchSize in tests/unit/commands/test_perf_cli.py covers the helper (dynamic / static / scalar / fallback), throughput scaling for both static-batch (regression guard: 100 not 800 sps) and dynamic-batch cases, the warning on mismatch, and the JSON field. Full test_perf_cli.py, test_perf_module.py, and test_perf_composite.py suites pass.

Scope note

This addresses the batch-size correctness aspect of the epic. The broader items (batch-size sweep --batch-sizes 1,4,8,16 and --ep all cross-EP comparison) are not part of this PR.

Closes #155

[xieofxie]

For the issue 155, only the remaining --ep all is tracked via #449, so close it

[DingmaomaoBJTU]

Overall this is a clean, well-motivated fix. The core bug (samples_per_sec inflated by requested batch on static-batch models) is correctly diagnosed and fixed. The effective_batch_size helper is clear and its fallback behavior is well-documented. Tests cover the important scenarios including the regression guard. A few minor observations below.