Summary

This PR fixes excessive memory allocation for Transformer attention ops when the
sequence length is statically known at compile time (e.g. seq_len=128).

For static-shape attention, Metal may eagerly allocate large intermediate buffers
(e.g. QKT matrices), which can lead to multi-GB allocations and OOM on iOS devices.
The existing attention slicing pass was gated behind a high sequence-length
threshold and did not trigger for smaller static shapes.

This change enables memory-efficient attention slicing for static sequence lengths
while preserving the existing behavior for dynamic-shape models.

Problem

When exporting Transformer models with a statically known sequence length,
scaled_dot_product_attention may materialize large intermediate tensors during
lowering. On iOS, this can result in excessive Metal buffer allocation (observed
~10GB) and OOM during inference or benchmarking, even for relatively small models
(e.g. Llama-style models with seq_len=128).

Solution

• Detect whether the attention sequence length is statically known.
• For static shapes with sequence length ≥ 64, apply the existing
  scaled_dot_product_attention_sliced_q pass to break the computation into
  smaller chunks and reduce peak memory usage.
• Preserve the original slicing threshold (1280) and behavior for
  dynamic-shape models to avoid unnecessary overhead.

This approach limits the change to the pathological static-shape case and avoids
global behavior changes.

Testing

• Added a regression test that constructs a Transformer-style attention block with
  a static sequence length (seq_len=128).
• The test verifies:
  • The sequence length is statically known.
  • The attention op is replaced by sliced logic.
  • Intermediate tensor sizes remain below a conservative safety bound, preventing
    pathological buffer materialization.
• Tests run at conversion time only and do not require iOS or Metal execution.

Notes

• This fix targets mobile inference scenarios where static shapes can trigger
  eager buffer allocation.
• Dynamic-shape models are intentionally unaffected.

Fixes #2590.