feat: create a tf2fhe pipeline

[asraa]

Blocked by #268

[j2kun]

Putting some notes here as I think about them:

• Start with a tosa model, where the model is a single function and all inputs are private.
• Run secretize to annotate the function args to be converted to secret
• Run wrap-generic to wrap the body of the annotated function in a secret.generic
• Run distribute-generic with option distribute-through=affine.for to distribute generic down to the bodies of loop nests.
• Run yosys-optimizer to optimize the individual generic bodies, resulting in an IR composed of secret.generic blocks (possibly inside loop nests) with booleanized comb ops inside and some secret.cast ops around each generic.
• Run comb-to-cggi which converts the comb ops and also handles the remaining removal of secret.generic, secret.cast, and secret.secret types.
• Run cggi-to-tfhe-rust and emit-tfhe-rust

[asraa]

Run yosys-optimizer to optimize the individual generic bodies, resulting in an IR composed of secret.generic blocks (possibly inside loop nests) with booleanized comb ops inside and some secret.cast ops around each generic.

After #308 the for loop bodies will contain affine.load / affine.store - which translateToVerilog will be able to handle - then after YosysOptimizer booleanizes the inputs, we'll be converting secret<memref<?xi8>> multi-bit types to secret<memref<tensor<?xi8>>> types.

Question: At what point is more accurate to call these memref<?xsecret<i8>>...

[j2kun]

Question: At what point is more accurate to call these memref<?xsecret<i8>>...

I think the types will be, in order of passes

• secret<memref<?xi8>>
• secret<memref<?xtensor<8xi1>>

It sounds like we'd want another bufferization pass... but that seems like it won't work if it starts from a mixed tensor/memref IR. I wonder if it makes sense to put bufferization later in the pipeline, after Yosys optimization? Then the types are tensor<?xtensor<8xi1>> and it should be easier to collapse that into a single multidimensional tensor.

[j2kun]

Some nits, otherwise LGTM

[asraa]

@j2kun pipeline cleaned up!