Use _weight_int8pack_mm for CPU + eager

[malfet]

Restrict use of the op to eager + torch > 2.3 (i.e. only nightlies) and cpu

This improves the perf for

% python3 torchchat.py generate --dtype float16 --device cpu --quant '{"linear:int8" : {}}' --checkpoint-path checkpoints/stories110M/model.pth

from 21 to 54 tokens/sec on M1 Max

[mikekgfb]

Is that because we are running with an old pytorch for Macos-12 ?
(if so we need to suppress this test on macos-12/x86 for now (but not on macos-14) cc: @guangy10

https://github.com/pytorch/torchchat/actions/runs/8824786469/job/24227882131?pr=472

  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/_ops.py", line 854, in __call__
    return self_._op(*args, **(kwargs or {}))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: _weight_int8pack_mm_cpu : expect A to be bfloat16 tensor.

[malfet]

@mikekgfb sure, I can spend some cycles doing it, but I wonder what's the goal. Are TorchChat claims to support Pytorch-2.2?)