[Dev] Bug fix for Block Reduce Template and improve TL

[LeiWang1999]

Fix bugs for incorrect thread binding and assignment for K Tile Size.

Improve TL Pass MergeStaticSharedMemory to Handle complex Memory Planing (For example, Stream-K).

Enhance LowerThreadAllReduce Pass to Auto Select Memory Scope based on the IR Script.

[LeiWang1999]

Also enhance auto tune.

@autotune(configs=get_configs(), keys=['block_row_warps', 'block_col_warps', 'warp_rows', 'warp_cols', 'chunk', 'reduce_k', 'num_stages', 'splitk_factor'], warmup=5, rep=20)
    @jit(out_idx=[2], supply_type=tl.TensorSupplyType.Normal, ref_prog=ref_program, rtol=1, skip_check=True, profiler='tvm')

Now we have profiler option ['tvm', 'torch'] to enable accurate kernel profile with TVM Runtime.

Also introduce tqdm to reveal the tune proc.