Add general convolution TTIR op. Add pass that legalizes it to Conv2D

[LPanosTT]

• Added ConvolutionOp, a very flexible convolution operation that can represent a convolution (and conv-transpose) of any dimensionality
  • Adding because stablehlo only has one convolution operation exactly like this, and we don't want to legalize to specific convolutions in dialect conversion, rather transform it within the same dialect.
  • Created a TTIR_ConvLayoutAttr to store the dimension numbers of the input, weight, and output.
    • I.e which index is the batch dimension, which index is the channels dimension, which indices are the spatial dimensions
• Added a pass that recognizes when a ConvolutionOp is a legal Conv2d (or can be converted to one by permuting the input, weights, and/or output) and performs the transformation.
• In the future we can add passes that legalize ConvolutionOp to any ConvNd or ConvNdTranspose.

[nsmithtt]

Sorry I missed this last time around, let's just make this a SmallVector<bool> boolArray(convolutionLayout.getInputSpatialDimensions().size()).

[LPanosTT]

I was able to use SmallVector instead of std::vector for all the attributes actually. Also was able to get rid of the helper because of that. Take a look :)

[nsmithtt]

Looks great!

[sdjordjevicTT]

This converts broad TTIR Conv to a more specific and narrow one. Should we leverage Jovan's work to consolidate the decomposition into a narrower set of TTIR operations?
#969

[LPanosTT]

This converts broad TTIR Conv to a more specific and narrow one. Should we leverage Jovan's work to consolidate the decomposition into a narrower set of TTIR operations? #969

Sure I suppose it can go in that pipeline. I'll put it in there if his PR gets merged first, otherwise I'll do it after.

[sdjordjevicTT]

I belive @jserbedzijaTT will merge this soon, can we wait a day or two to consolidate things from the start? :)