Adding JIT cache

[bringlein]

The launch overhead of triton kernels is a well known problem (see e.g. 1, 2, 3). Parts of the launch overhead comes from the fact that the triton JIT checks very carefully if an existing binary is safe to use.

In many scenarios, these checks can be relaxed.
This PR adds such a cache with relaxed checks is implemented by triton_dejavu.jitcache. It is implemented as a decorator that could be used in front of the triton.jit decorator:

@triton_dejavu.jitcache(
    check_keys=["x", "BLOCK_SIZE", "USE_ALIBI_SLOPES", "SLIDING_WINDOW", "filter_by_query_len"],
)
@triton.jit
def kernel_paged_attention_.... 

Details see Readme.