ToRadixBE: polish error handling and align with brillig VM and AVM circuit

[jeanmon]

Discrepancy with BrilligVM:

• if radix > 256: no error is handled by brilligVM
• if num_limbs < 1: no error is handled by brilligVM

Both BrilligVM and AVM TS simulator:

• no explicit handling of error if radix == 0
• no enforcement that radix == 2 whenever bit mode is activated (TS would hit an assert if radix > 2 with bit mode activated)

Note on radix ==1: Current code would return zero values which I think it is perfectly fine. Should we enforce radix > 1

Apply the above to the AVM circuit (possibly unconstrained).

[jeanmon]

Feedback from @guipublic on how ACIR toradix handles these edge cases:

• If radix == 0, returns an error
• if radix > 256, returns an error
• If num_limbs ==0, valid only if the input == 0 otherwise return an error

Remark 1: ACIR additionally restricts radix to be a power of two (Brillig VM is more permissive)
Remark 2: ACIR toradix does not support any input related to bit vs byte mode

@guipublic Do you mind mentioning how radix == 1 is handled here?

[guipublic]

Noir ACIR handling of to_radix(x,n,r), where x is the input, n is the number of limbs, and r is the radix

limbs nb 0: generate x = 0, a constraint which works only if input is 0.
radix 0: crash during compilation, before getting the panic message radix must be a power of 2
radix 1: 'properly' constrain the limbs to be 0 bits, but execution always fail. Not sure if BB properly handles 0 bits range constraints.
radix 5: ICE: Radix must be a power of 2
radix > 256: Not supported in Noir (limbs are represented as bytes), but there is no (more) check. It compiles and execute successfully. For instance a to_radix of x=300 with radix r=512 will give the limbs [44].

So we have at least 3 bugs, for radix 0, 1, and >256.
I think Noir should enforce radix to be a power of 2 between 2 <= r <=256.
I created an issue noir-lang/noir#7117