[samansg-aws]

Issue #, if available: n/a

Description of changes:

SelectiveAC hardcodes "recompute every second matmul" (meta[mm_count_key] % 2 == 0),
which is the only memory/compute tradeoff per-op SAC exposes.
force_recompute_mm_shapes_by_fqns can only push more matmuls into
recomputation, never fewer, so there is no way to say "save all the matmuls".

Adds one field to SelectiveAC.Config:

• matmul_recompute_freq: int = 2 -- recompute one matmul out of every n.
• 2 (default) reduces to the old predicate, so existing configs are bit-identical.
• 1 recomputes every matmul; 0 saves every matmul; negative raises ValueError.
• Matmuls matched by force_recompute_mm_shapes_by_fqns are still always
  recomputed, so the two knobs compose as before.

Tests in tests/unit_tests/gpu/test_activation_checkpoint.py count aten.mm
calls per weight during backward (same technique as test_force_recompute_mm_fqns)
for freq in {0, 1, 2, 3}, plus the ValueError case.

pre-commit run --files passes on both changed files.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.