Refactor throughput dtype inference into Precision plugin API

[KAVYANSHTYAGI]

Summary

This PR moves throughput dtype inference into the Precision plugin itself by introducing a new method:

compute_dtype()

Instead of using isinstance(...) checks inside throughput utilities, we now simply call:

plugin.compute_dtype()

What Changed

• Added compute_dtype() to the base Precision class (default: torch.float32)
• Implemented compute_dtype() across all built-in Fabric and PyTorch precision plugins
• Removed hard-coded type checks and internal attribute access in throughput utilities
• Simplified _plugin_to_compute_dtype() to delegate directly to the plugin API
• Added tests to verify correct behavior and integration with ThroughputMonitor

Why

Previously, throughput relied on chained isinstance(...) checks and accessed internal plugin attributes.
This tightly coupled throughput logic to specific precision implementations.

By moving dtype responsibility into the plugin:

• We improve encapsulation
• Reduce maintenance overhead
• Automatically support custom precision plugins
• Keep throughput utilities clean and extensible

fix #21535

[codecov]

Codecov Report

❌ Patch coverage is 84.74576% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 87%. Comparing base (7d2de87) to head (93dcbd5).
⚠️ Report is 1 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #21537   +/-   ##
=======================================
  Coverage      87%      87%           
=======================================
  Files         270      264    -6     
  Lines       24071    23980   -91     
=======================================
- Hits        20867    20861    -6     
+ Misses       3204     3119   -85