[Tosa] : Use output type for bias for creating tosa.conv

[sahas3]

For ConvolutionLayer initialized without bias, a zero tensor for bias is created when converting to tosa.conv2d as the op always expects a bias tensor. This zero tensor was always initialized to be fp32 irrespective of what the input/weights type were. This leads to a validation error since bias type (fp32) didn't match with output of conv (fp16) when the input/weight are of fp16 type.

[sahas3]

The nightly build failures are unrelated and seems to be failing for all recent CIs. @sjarus, please let me know if there's any concern with merging this change after approval. Thanks!

[sahas3]

Converting to draft as I found some other failures when trying different integer type inputs, will look into those failures and possible add fixes in this same PR.

[sjarus]

Nice catch, thanks @sahas3 !

[sahas3]

Nice catch, thanks @sahas3 !

Thanks @sjarus. However, I had to make some changes on top of what you had already reviewed to handle some edge cases captured in the changes in the basic.mlir file. Sorry for the inconvenience, but please can you take another look when you have a chance? Thanks!

[sjarus]

This is quite extensive in terms of introducing new capability around quantized type coverage. Thanks so much.

cc: @eric-k256 @justin-ngo-arm

[sahas3]

This is quite extensive in terms of introducing new capability around quantized type coverage. Thanks so much.

Thank you, appreciate the kind words.