[CuTeDSL] Make autotuning more user friendly & new docs

[kainzhong]

This PR adds a lot of things to the overall CuTeDSL autotuning process to make it easier to use and more convenient:

• Add derived_params to autotuning API so it computes derived arguments that must be decided during compile time but rely on autotuned parameters (whose values are different per configuration tuned)
• Make testing.autotune_jit able to decorate a class. This is for people who prefer to implement kernels as classes that fix the compile time params in __init__ method and pass runtime params to __call__, so now they can do this

@testing.autotune_jit(
    params_dict={"COPY_BITS": [32, 64, 128], "THREADS_M": [4, 8]},
    update_on_change=["M", "N"],
    warmup_iterations=10,
    iterations=50,
)
class ElementwiseAddKernel:
    THREADS_M = 4

    def __init__(self, COPY_BITS=128):
        self.COPY_BITS = COPY_BITS

    @cute.jit
    def __call__(self, mA, mB, mC, M, N):
        elementwise_add(
            mA, mB, mC, M, N,
            copy_bits=self.COPY_BITS,
            threads_m=self.THREADS_M,
        )

• Support pruning undesired configurations (either invalid or known to be suboptimal) -- this aligns with triton's prune_configs_by in their autotuning API
• Add autotune_suite API so now user can provide a list of representative shapes and find the best overall kernel candidates with the ranking rule: sort according to how many times a configuration is accepted across all inputs, where "accepted" mean for a input case the configuration is in the top accept_percentile fastest ones
• Better autotuning docs. Previously the doc only explains how to autotune manually which is quite painful at least for me