llama: use a more efficient rope implementation

[rattus128]

Get rid of the cat and unary negation and inplace add-cmul the two halves of the rope. Precompute -sin once at the start of the model rather than every transformer block.

This is slightly faster on both GPU and CPU bound setups.

I found this profiling ace-step dynamic_vram and the RoPE was the biggest thing on the flame graph.

Example test conditions:

RTX5090, linux, Ryzen 9600 underclocked to 2GHz (CPU bound)
Ace step 1.5 turbo 195s

Before:

LM sampling: 100%|██████████| 975/975 [00:15<00:00, 63.05it/s]
Requested to load ACEStep15
loaded completely; 25390.26 MB usable, 4565.35 MB loaded, full load: True
100%|██████████| 8/8 [00:00<00:00,  9.51it/s]
Requested to load AudioOobleckVAE
loaded completely;  321.70 MB loaded, full load: True
Prompt executed in 30.84 seconds

After:

LM sampling: 100%|██████████| 975/975 [00:14<00:00, 66.60it/s]
Requested to load ACEStep15
loaded completely; 25390.26 MB usable, 4565.35 MB loaded, full load: True
100%|██████████| 8/8 [00:00<00:00,  9.56it/s]
Requested to load AudioOobleckVAE
loaded completely;  321.70 MB loaded, full load: True
Prompt executed in 30.02 seconds