Some Conv tests failing with BF16

[pranavm-nvidia]

Some of the convolution integration tests are failing with BF16.

There are two types of failures:

1. allclose fails:

>       assert tp.allclose(tripy_out, torch_out)
E       assert False
E        +  where False = <function allclose at 0x76b9495f2280>(tensor(\n    [[[[338]],\n\n      [[880]]]], \n    dtype=bfloat16, loc=gpu:0, shape=(1, 2, 1, 1)), tensor(\n    [[[[338]],\n\n      [[880]]]], \n    dtype=bfloat16, loc=gpu:0, shape=(1, 2, 1, 1)))
E        +    where <function allclose at 0x76b9495f2280> = tp.allclose

Note that the Tripy/torch outputs compare equal if they are evaluated separately. It's only when the allclose is part of the same computation graph as the Conv that it returns a False output. This implies a bug in a graph transformation.

2. Myelin error:

E       MTRTException: failed to run pass pipeline
E           Internal Error: MyelinCheckException: conv_lowering.cpp:43: CHECK(op->operands()[0]->is_tensor()) failed. 
E           Error Code: 9: Skipping tactic 0x0000000000000000 due to exception [myelin_graph.h:attachExceptionMsgToGraph:840] MyelinCheckException: conv_lowering.cpp:43: CHECK(op->operands()[0]->is_tensor()) failed. 
E           IBuilder::buildSerializedNetwork: Error Code 10: Internal Error (Could not find any implementation for node {ForeignNode[[tensorrt.constant] (%t3288) ...[tensorrt.cast] (%t3288) ]}.)
E           (%t3288) error: failed to translate function 'tensorrt_cluster' to a TensorRT engine

Most likely, it's expecting a build-time constant here. Constant folding may solve this.