Granite FP8 (E4M3) model enablement [WIP]

[rishirajgupta-2095]

Runs compressed-tensors FP8 (E4M3) Granite checkpoints in FP8 on Spyre via
spyre.scaled_mm, instead of dequantizing them to 16-bit at load.

Model: ibm-granite/granite-3.3-8b-instruct-FP8
(baseline: ibm-granite/granite-3.3-8b-instruct)

Related: torch-spyre/torch-spyre#4432

Initial Results

Granite 3.3 8B on current main, prefill S=512, 200 of 280 projections in FP8:

	base (bf16)	FP8
steady-state decode	268.1 ms/token	225.4 ms/token (−16%)
load	37.8 s	62.7 s
weight HBM	~16 GB	~10.7 GB

Requirements

torch-spyre with eager FP8 ops (torch-spyre/torch-spyre#3172)

Known limitations

• o_proj and down_proj stay fp16 pending torch-spyre FP8 codegen fixes.

• RMSNorms feeding FP8 projections use an fp16 variance reduction on Spyre:
  stock RMSNorm's fp32->fp16 cast leaves torch-spyre no feasible layout for the
  FP8 scaled_mm. Bound per instance, FP8 checkpoints only; TODO to drop.

  Testing

  pytest -s -vvv tests/spyre/test_e2e_smoke_spyre.py
  --model-path ibm-granite/granite-3.3-8b-instruct-FP8
  pytest -s -vvv tests/spyre/test_e2e_smoke_spyre.py
  --model-path ibm-granite/granite-3.3-8b-instruct
  PYTHONPATH=. python scripts/run_fp8_granite_spyre.py