Remove missed out .float() cast and add dtype-safe handling for torch.linalg.inv in transfuser model

[kamalrajkannan78]

Ticket

• closes Enable bfloat16 support in transfuser model tt-xla#3196

Problem description

• This PR missed removing a .float() cast, which causes the entire post-processing part of the model to run in fp32.

What's changed

• .float() changed to .to(scores.dtype): Now topk_xs inherits the dtype from the input scores tensor
• torch.linalg.inv doesn't support Low precision dtypes. so the input is cast to fp32 before the operation and the output is cast back to the original dtype.

Checklist

• Verified the changes through local testing

Logs

CPU runs

• FP32 (Before vs. After Fix): Results remain unchanged after introducing this PR changes - https://www.diffchecker.com/2hodEVYa/
  • feb11_transfuser_fp32_cpu_before_fix.log
  • feb11_transfuser_fp32_cpu_after_fix.log
• BFP16 - feb11_transfuser_cpu_bfp16.log

TT-XLA runs

• xla_run.zip