[AutoDeploy] Expose logit_softcap parameter in flashinfer_attention

Flashinfer supports logit_softcap in attention. However, we currently do not expose this parameter in the flashinfer attention interface. We should expose so we can later set it as needed.

More detailed description:

• Just like scale is a constant in the FlashInfer Attention Backend in tensorrt_llm/_torch/auto_deploy/custom_ops/flashinfer_attention, we also want logit_softcap from flashinfer to be a layer-specific constant we can pass through the attention interface.
• Let's expose it but let's set the default to whatever the default is in the flashinfer op so it becomes optional
• Let's also modify the corresponding unit test in the folder tests/unittest/_torch/auto_deploy to test for this new use case / parameter