fix(dynamo): make truncate_double metadata match engine precision

[narendasan]

Summary

• use partition FX metadata instead of executing TensorRT subgraphs during compilation
• record truncated float32 engine bindings accurately
• preserve the explicit float64 restoring cast and its output metadata
• handle non-tensor outputs and CPU constants without probing the engine

Issues

Closes #4604
Closes #4623
Closes #4624

Tests

• pytest tests/py/dynamo/conversion/test_truncate_double.py -n0
• original reproducers for 🐛 [Bug] repair_double_inputs runs the TRT subgraph to read metadata, crashing on non-tensor outputs and CPU constants #4604, 🐛 [Bug] truncate_double inserts a float64 restoring cast but leaves the engine node's metadata claiming float64 #4623, and 🐛 [Bug] engine meta kernel reports the pre-truncation output dtype (float64) while the engine returns float32 #4624 no longer reproduce

Stack

4 of 7. Base: #4629. Follow-up: narendasan/fix-dynamic-scatter-index-put.