Change default sliding window pattern to the recommended "L" when FA3 is not available

[ddudek]

Changes default setting of sliding window pattern for setups without out-of-the-box FA3 support to the recommended in the warning.

This simplifies configuration for beginners running nanochat on their local setups, e.g. consumer grade GPUs like 3090/4090 and others without FA3 support.

Before:

$ python -m scripts.base_train --depth=12 --device-batch-size=16
...
GPU: NVIDIA GeForce RTX 3090 | Peak FLOPS (BF16): 7.10e+13
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
WARNING: Flash Attention 3 not available, using PyTorch SDPA fallback
WARNING: Training will be less efficient without FA3
WARNING: SDPA has no support for sliding window attention (window_pattern='SSSL'). Your GPU utilization will be terrible.
WARNING: Recommend using --window-pattern L for full context attention without alternating sliding window patterns.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Vocab size: 32,768
Model config:
{
  "sequence_len": 2048,
  "vocab_size": 32768,
  "n_layer": 12,
  "n_head": 6,
  "n_kv_head": 6,
  "n_embd": 768,
  "window_pattern": "SSSL"
}
...
step 00011/02205 (0.50%) | loss: 8.170549 | lrm: 1.00 | dt: 13119.33ms | tok/sec: 39,963 | mfu: 45.15 | epoch: 1 | total time: 0.22m | eta: 479.7m

After:

$ python -m scripts.base_train --depth=12 --device-batch-size=16
...
GPU: NVIDIA GeForce RTX 3090 | Peak FLOPS (BF16): 7.10e+13
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
WARNING: Flash Attention 3 not available, using PyTorch SDPA fallback
WARNING: Training will be less efficient without FA3
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Vocab size: 32,768
Model config:
{
  "sequence_len": 2048,
  "vocab_size": 32768,
  "n_layer": 12,
  "n_head": 6,
  "n_kv_head": 6,
  "n_embd": 768,
  "window_pattern": "L"
}
...
step 00011/02205 (0.50%) | loss: 8.177470 | lrm: 1.00 | dt: 7127.85ms | tok/sec: 73,554 | mfu: 91.90 | epoch: 1 | total time: 0.12m | eta: 260.6m

[svlandeg]

Nice idea and super minimal edit. As you mention, merging this would make the script behave more user-friendly for beginners with non-FA3 setups.