[Gluon][Tutorial] Merge d64 and d128 attn kernels (#7226)

This PR deletes the d64 kernel and only uses the d128 kernel by enabling
BLOCK_N=128 for d64 and increasing the number of load buffers, as well
as reordering code in the softmax partition to improve performance.

This increases the peak performance up to 770 tflops (780 for
non-causal) for DHEAD=64. However, this makes performance for smaller
ctx (16k and below) worse than before because the scheduling granularity
is worse. This should get fixed when the kernel is made persistent, and
this means I don't have to optimize two separate kernels.

DHEAD=128 performance is unchanged.

```
Attention Z=4 H=32 D=64 causal=False:
     N_CTX  triton-fp16  cudnn-fp16
0   1024.0   381.658388  480.100511
1   2048.0   569.136004  804.057444
2   4096.0   663.284730  892.582773
3   8192.0   737.043055  939.454979
4  16384.0   764.685874  952.196135
5  32768.0   778.020934  931.158252
6  65536.0   785.854946  912.826103
Attention Z=4 H=32 D=64 causal=True:
     N_CTX  triton-fp16  cudnn-fp16
0   1024.0   167.438219  351.715759
1   2048.0   295.866693  619.581918
2   4096.0   452.670120  754.473029
3   8192.0   586.122336  833.803614
4  16384.0   659.646168  831.281960
5  32768.0   743.090000  910.116397
6  65536.0   771.990185  918.770994
Attention Z=4 H=32 D=128 causal=False:
     N_CTX  triton-fp16   cudnn-fp16
0   1024.0   629.445279   915.936310
1   2048.0   930.445757  1222.156019
2   4096.0   984.129646  1306.298210
3   8192.0  1107.853967  1382.667255
4  16384.0  1175.943239  1265.863845
5  32768.0  1187.172469  1248.898664
6  65536.0  1174.481087  1268.167581
Attention Z=4 H=32 D=128 causal=True:
     N_CTX  triton-fp16   cudnn-fp16
0   1024.0   265.095023   547.764710
1   2048.0   472.862887   852.318204
2   4096.0   690.310855  1085.992183
3   8192.0   882.526991  1247.025302
4  16384.0  1062.367442  1292.686167
5  32768.0  1131.631485  1196.174915
6  65536.0  1122.032238  1248.103686
```

Author: Mogball