fix: force float32 precision on CPU to prevent distorted structures

[joshuasteier]

This updates PR #670 from the blanket CPU float32 workaround in main.py
to the underlying fix described in #662.

On CPU, PyTorch Lightning wraps the forward pass with
torch.autocast("cpu", dtype=torch.bfloat16), so
torch.autocast("cuda", enabled=False) is a no-op there. As a result,
the structure and affinity blocks in boltz2.py still run under CPU bf16
autocast even though they explicitly cast inputs to float32.

This patch:

• removes the main.py CPU precision workaround from this PR
• adds an autocast_device_type() helper in model/modules/utils.py
• replaces the three hardcoded torch.autocast("cuda", enabled=False)
  blocks in src/boltz/model/models/boltz2.py with
  torch.autocast(autocast_device_type(s.device.type), enabled=False)

Local validation:

• under torch.autocast("cpu", dtype=torch.bfloat16), the old path keeps
  matmul in torch.bfloat16
• the new autocast_device_type(...) path correctly yields torch.float32

Credit:

• @Christopher-Lee-McClendon for the root-cause diagnosis in CPU inference produces lower quality predictions due to device-specific autocast bug in boltz2.py #662
• @volgin for the helper-based fix pattern used in boltz-community
• @ozvoz for validating the earlier workaround in CPU runs produce distorted structures with excessively long bonds and colliding atoms #653

Closes #653
Closes #662