Add TensorRT `IErrorRecorder` Implementation

[maedtb]

• Adds a new thread-safe python class TrTErrorRecorder which implements the TensorRT IErrorRecorder interface. This class captures errors to display to the user, and can optionally terminate TensorRT processing when errors occur.
• We now set the error_recorder field on the TensorRT tensorrt.Builder and tensorrt.Runtime classes to an instance of TrTErrorRecorder.
• We now check for errors while initializing TensorRT engines, raising exceptions if TensorRT reports any errors to us.

[maedtb]

I ran into several errors using TensorRT, and it was unclear why these errors were happening. Adding the TensorRT error reporter cleared up the issues for me entirely. The error messages TensorRT report to us look something like this:

[defaultAllocator.cpp::allocate::31] Error Code 1: Cuda Runtime (out of memory)
[executionContext.cpp::ExecutionContext::565] Error Code 2: OutOfMemory (Requested size was 30152807424 bytes.)

It's not the most user friendly, but it's a lot more helpful of an error than a 'NoneType' object has no attribute 'set_input_shape' :^)

[mcmonkey4eva]

Tested this - it works properly for converting and generating, but an error during generation while it logs properly doesn't raise an exception

so it just generates a black image while spamming console with a new error every step

[comfyanonymous]

[07/15/2024-23:30:09] [TRT] [I] [MemUsageStats] Peak memory usage during Engine building and serialization: CPU: 13075 MiB
[07/15/2024-23:30:09] [TRT] [I] Serialized 16605 bytes of code generator cache.
[07/15/2024-23:30:09] [TRT] [I] Serialized 4704604 bytes of compilation cache.
[07/15/2024-23:30:09] [TRT] [I] Serialized 2218 timing cache entries
Segmentation fault

This PR gives me a segfault when I try to convert SD1.5

[maedtb]

Just leaving an note here that I'm planning on investigating these issues this weekend and I haven't abandoned this. Going to pull it into Draft in the mean time.