Fix landmark attention patch

[NanoCode012]

This updates the monkey patch method, fixes inference.

• Test

[NanoCode012]

Running 8bit

n=5k

Running test 10/10 for n = 5000...
Number of tokens in this prompt:  1405
Expected number in the prompt: 15641, mem output: 15641
Success!
Accuracy mem for n = 5000: 100.0%

n=8k

Running test 10/10 for n = 8000...
Number of tokens in this prompt:  2205
Expected number in the prompt: 45997, mem output: 45997
Success!
Accuracy mem for n = 8000: 80.0%

n=10k

Running test 10/10 for n = 10000...
Number of tokens in this prompt:  2738
Expected number in the prompt: 11032, mem output: 11032
Success!
Accuracy mem for n = 10000: 80.0%

n=12k

Running test 10/10 for n = 12000...
Number of tokens in this prompt:  3272
Expected number in the prompt: 48243, mem output: 48243
Success!
Accuracy mem for n = 12000: 80.0%

n=14k

Running test 10/10 for n = 14000...
Number of tokens in this prompt:  3806
Expected number in the prompt: 25702, mem output: 25702
Success!
Accuracy mem for n = 14000: 100.0%

n=18k

Running test 10/10 for n = 18000...
Number of tokens in this prompt:  4872
Expected number in the prompt: 16625, mem output: 16625
Success!
Accuracy mem for n = 18000: 80.0%

n=20k

Runnin
Running test 10/10 for n = 20000...
Number of tokens in this prompt:  5405
Expected number in the prompt: 23287, mem output: 23287
Success!
Accuracy mem for n = 20000: 80.0%

n=25k

Running test 10/10 for n = 25000...
Number of tokens in this prompt:  6738
Expected number in the prompt: 25475, mem output: 25475
Success!
Accuracy mem for n = 25000: 80.0%

n=38k

Running test 10/10 for n = 38000...
Number of tokens in this prompt:  10206
Expected number in the prompt: 48851, mem output: 48888
Fail.
Accuracy mem for n = 38000: 40.0%