[AIROCMLIR-707] Make attentionSweeps split_kv filter device-memory-aware

[bogdan-petkovic]

Motivation

Some attention sweep samples with large split_kv generate very high temporary storage pressure and are currently rejected as FAIL after entering the pipeline.
This PR adds an early sweep-side prefilter so memory-heavy splitKV cases can be skipped before expensive execution, reducing wasted sweep effort while keeping existing compiler/runtime behavior unchanged.

Technical Details

Updated mlir/utils/performance/attentionSweeps.py to add a device-memory-aware splitKV prefilter in sampling:

• Added splitKV extra-storage estimator for sampled attention shapes.
• Added default splitKV limit policy based on visible device memory:
  • deviceMem / 8, clamped to [1 GiB, 8 GiB]
  • fallback to 1.5 GiB when memory query is unavailable.
• Added CLI control:
  • --splitkv-extra-bytes-limit (non-negative int64-validated override).
    Refactored sample filtering to track reasons separately:
• MAX_TOKENS filter
• splitKV extra-storage filter
• cumulative reporting across initial and refill sampling batches.
  Added focused unit tests in mlir/utils/performance/tests/test_attentionSweeps.py:
• limit policy behavior (clamp/fallback),
• splitKV estimator behavior,
• per-reason filter accounting.
  No compiler/verifier logic was changed in this PR; scope is limited to sweep generation/filtering behavior.

Test Plan

• Run attentionSweeps.py through CI to validate end-to-end sweep behavior with the new splitKV prefilter.
• Confirm that memory-heavy splitKV cases are filtered before expensive execution stages.

Test Result

• CI attention-sweep run

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.