[Example] Add example for decoding kernel of flash linear attention

[yaoyaoding]

This PR adds the implementation for the decoding kernel of flash linear attention:
https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/fused_recurrent.py

      name      (B, T, H, K)   (HV, V) latency (ms)
0   triton    (1, 1, 4, 128)  (8, 128)        0.007
1    tilus    (1, 1, 4, 128)  (8, 128)        0.006
2   triton    (1, 2, 4, 128)  (8, 128)        0.008
3    tilus    (1, 2, 4, 128)  (8, 128)        0.007
4   triton    (1, 4, 4, 128)  (8, 128)        0.011
5    tilus    (1, 4, 4, 128)  (8, 128)        0.009
6   triton    (1, 8, 4, 128)  (8, 128)        0.016
7    tilus    (1, 8, 4, 128)  (8, 128)        0.013
8   triton   (1, 16, 4, 128)  (8, 128)        0.026
9    tilus   (1, 16, 4, 128)  (8, 128)        0.022
10  triton   (1, 32, 4, 128)  (8, 128)        0.047
11   tilus   (1, 32, 4, 128)  (8, 128)        0.044
12  triton   (1, 64, 4, 128)  (8, 128)        0.081
13   tilus   (1, 64, 4, 128)  (8, 128)        0.086
14  triton  (1, 128, 4, 128)  (8, 128)        0.153
15   tilus  (1, 128, 4, 128)  (8, 128)        0.161

To support this kernel, have some other enhancements:

1. add sqrt operator
2. move the layout inference verification to have better visualization in debug log