Improve compute op coverage rate

[shengfukevin]

et_replay skipped some operators, need to understand the reason why these operators are skipped, and how to fix it.

Some known ops:

1. _scaled_dot_product_efficient_attention_backward_cuda (op schema does not support optional attention_bias)

2. fbgemm for embedding table
   The current implementation tries to recover the embedding table lookup op from its forward call and backward call (The way to match the forward call and backward call is also fragil). It is not reliable as the op itself keep changing.