Fix model breakages

[kevinwuTT]

Model status for this PR:

⚠️ GPT-2 fixes

Unmark gpt2 and mnist test models to expect passing
Disable conversion from aten._to_copy
Pass device for all from_torch ops Reverted because this conflicts with unsqueeze conversion
Replace aten.full op to a literal scalar for certain cases
Compare only Tensor types for dictionary outputs
Fails because:

• Explicit casting of dtype to ttnn.bfloat16 for from_torch op makes GPT-2 fail. Removing it causes, Bloom, Llama, and Yolos to fail.

Failed to generate binaries for embeddings_tilize TT_THROW @ ../tt_metal/jit_build/build.cpp:396: tt::exception
info:
ncrisc build failed

MNIST fixes:

Replace aten.view with aten.reshape

⚠️ Falcon-7B status:

Fails because:

• layer_norm conversion

TT_THROW @ ../tt_metal/impl/program/program.cpp:489: tt::exception
info:
Statically allocated circular buffers on core range [(x=0,y=0) - (x=0,y=0)] grow to 1611072 B which is beyond max L1 size of 1048576 B

Bloom and Llama fixes:

Add conversion for aten.min
Add exception to aten.eq conversion
Fix reusing ttnn data movement op if mixed with aten ops
Convert all inputs to ttnn.bfloat16 when moving data in
Skip unsqueeze transformation if last dim of input is not the same as…
Add exception to aten.expand conversion when last dimension of input …
Support list type arguments
Check layout change for ttnn reshape and embedding op
Freeze encoder for llama model

Yolos fixes

Add workaround for ttnn.permute when dim 0 is 1 for rank 3
Reconvert int64 types from metadata when mixing ttnn and aten ops
Check for valid page size for ops that decompose to ttnn.full
Delete aten.expand op if output has the exact same shape

General fixes:

Consolidate metadata during op conversion
Fix output type of aten.arange unit test to match output of original
Disable to_copy unit test to re-evaluate conversion
Lower pcc for addmm slightly
Change input shapes of some unit test to match exceptions in current …