Implement remaining canonicity checks

[iansuvak]

Summary

This PR implements missing canonical checks for Marshall validate defined here:

https://github.com/algorandfoundation/specs/blob/master/dev/crypto.md

Specifically it implements conditions: 2-5 listed here:

1. Maps must omit key-value pairs where the value is a zero-value, unless otherwise specified;
2. Positive integer values must be encoded as "unsigned" in msgpack, regardless of whether the value space is semantically signed or unsigned;
3. Integer values must be represented in the shortest possible encoding;
4. Binary arrays must be represented using the "bin" format family (that is, use the most recent version of msgpack rather than the older msgpack version that had no "bin" family).

TODO

• Implement minimal length encoding checking for strings as well

Testing

Currently, the generated code compiles but isn't yet tested for correctness hence draft status.

[iansuvak]

This is not robust. If IfZeroExpr() returns empty string it will validate unconditionally

[zeldovich]

I think this should be i << 5, because msgpack supports only 5-bit negative fixint.

[zeldovich]

Actually, I think there might be two other issues here. First, it seems like you should be shifting right (i.e., i >> 5). Second, for signed values, I think you should check for == -1 rather than == 0.

[iansuvak]

Thanks for the catch, hopefully would have caught it soon in testing. JW how much of a performance vs readability tradeoff would it be to compare against int64 constants in this case rather than doing the bitshift comparisons?

[zeldovich]

I doubt there's a measurable performance difference.

Just out of curiosity, I tested go 1.21.0, and the compiler emits two instructions (CMPQ and SETCS) for x < 256, and emits three instructions (SHRQ, TESTQ, SETEQ) for (x >> 8) == 0. I didn't bother benchmarking them; we should go for what's clearer.

I don't have a strong opinion about whether bit-shifts or comparisons are clearer here.