Add int8 per channel weight-only quantized matmul

[lsy323]

Add the first xla quantized ops for per-channel weight-only quantized matmul.

The math is out[bf16] = matmul(act[bf16], weight[s8]) * scaler[bf16], the same as what was adopted in the XLA llama quant implementation

User experience:

• Call quantized op with already quantized weight in model code
• Swap the nn.Linear Module with the added quantized module in model code

More details about user experience can be found in the added README.

Changes:

• Added custom torch op and nn.Module for the quantized op
• Added user guide

Test:

• Test the lowered HLO is doing what we expect
• Test it works with Dynamo
• Test numerical correctness

int4 and blockwise quant support will be added in following PRs

[JackCaoG]

mostly lgtm, minor nits