Fusion rule for handling transformers exported models

[justinchuby]

When torch.onnx exports a model from transformers with SDPA, it generates a Concat
node to concatenate past_key/value with the new key/value to produce the graph output
for kv cache. This pattern can be fused into the Attention node, which has present_key
and present_value outputs. It is necessary for ONNX Runtime because it requires the outputs
to be produced by the Attention node when past_key and past_value inputs are provided.

[codecov]

Codecov Report

❌ Patch coverage is 73.68421% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 70.39%. Comparing base (811937c) to head (c408516).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...ipt/rewriter/rules/fusion/_attention_present_kv.py	76.47%	4 Missing ⚠️
onnxscript/rewriter/onnx_fusions/_onnx_fusions.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2632   +/-   ##
=======================================
  Coverage   70.38%   70.39%           
=======================================
  Files         222      223    +1     
  Lines       26288    26309   +21     
  Branches     2629     2629           
=======================================
+ Hits        18503    18519   +16     
- Misses       6865     6870    +5     
  Partials      920      920           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

This is still useful when enable_gqa=True