Add Salesforce xLAM model series

[zuxin666]

This PR add Salesforce/xLAM-7b-r, Salesforce/xLAM-8x7b-r, Salesforce/xLAM-8x22b-r models.

Note that the Salesforce/xLAM-8x7b-r model requires 8x40GB GPUs like A100 for inference, while the Salesforce/xLAM-8x22b-r model requires 8x80GB GPUs to run.

Tested these models with 8 H100 GPUs locally. The reference combined performance is as follows:

Here's the CSV table converted to markdown format:

Here's the CSV table converted to markdown format:

Rank	Overall Acc	Model	AST Summary	Exec Summary	Irrelevance Detection	Relevance Detection
1	87.31%	xLAM-8x22b-r (FC)	82.76%	92.39%	74.96%	97.56%
2	83.38%	xLAM-8x7b-r (FC)	78.18%	89.02%	72.35%	92.68%
3	80.33%	xLAM-7b-r (FC)	74.00%	85.43%	72.88%	92.68%
4	80.18%	xLAM-7b-fc-r (FC)	72.78%	87.68%	79.54%	80.49%
5	75.43%	xLAM-1b-fc-r (FC)	65.72%	83.30%	60.65%	97.56%

@HuanzhiMao @CharlieJCJ Would you please take a look? Thanks for the great benchmark!

[HuanzhiMao]

Thanks for the PR @zuxin666!

One clarification question: Is Salesforce/xLAM-7b-r an FC model or a prompting model?

The rest looks good to me.

[liuzuxin]

Hi @HuanzhiMao , they are all FC models that share the same prompting format, so they can also share the same xLAMHandler. Thanks!

[HuanzhiMao]

LGTM