Support hybrid trace and add benchmark

[yitongh]

• support hybrid trace
• register FA as PyTorch op
• add llm benchmark
• move patch to accelerate.apply_patch

[hanwen-sun]

It's a little strange to specify both eager/lazy backend and partial_compile/hybrid_trace in the command. Can we hide the backend option in the command(reserve in the Backendconfig)?
for example, --lazy represents LTC trace and lazy mode, --eager represents eager mode and cuda native, --hybridtrace represents lazy mode but hybridtrace, --partial_compile represents eager mode but dynamo + xla.
We only need to check if xla is aviliable in the environment for lazy mode.
What's more, we need add comments or docs for these options.