perf(custom_ops): compute FP8 per-token scale via quantscalepertokenfp8

[prashantgupta24]

Description

Replace the eager per-token FP8 activation scale computation in SpyreFp8LinearKernel with torch.ops.spyre.quantscalepertokenfp8. The op performs the amax reduction, scale division, and hardware clip inside its own compiled Inductor graph, avoiding an eager amax/division boundary and matching the deeptools operator contract. It is kept separate from the FP8 matmul graph because fusing it with _scaled_mm, quantize_fp8_with_scale, and quantize_weight_fp8_with_scale causes the SDSC compiler to fail on the combined kernel. Per-tensor activations keep the existing eager scalar scale path because quantscalepertokenfp8 always reduces over the hidden dimension.

Related Issues

Relates to the torch-spyre quantscalepertokenfp8 op introduced in torch-spyre/torch-spyre#3457.

Test Plan

• bash format.sh — passed
• uv run pytest tests/test_fp8_linear.py -m "not upstream" — 17 passed
• Microbenchmark on Spyre hardware comparing old eager-scale path to new in-graph path showed identical outputs for tested shapes.

Checklist

• I have read the contributing guidelines
• My code follows the project's code style (run bash format.sh)
• I have added tests for my changes (if applicable)
• I have updated the documentation (if applicable)
• My commits include a Signed-off-by: line (DCO compliance)

[github-actions]

👋 Hi! Thank you for contributing.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, run ./format.sh.
Now you are good to go 🚀.

We also recommend installing prek and configuring it to check your code before every local commit.