Remove truncation from Instruction::Cast to allow unsafe casts

[TomAFrench]

I've been doing some thinking about how we can perform "unsafe maths" to avoid all those nasty overflow checks and similar.

In general how this is being done is by converting to Field type, doing your operations and then converting back to the desired type. This runs into the trouble that converting back to the original type will automatically include an overflow check despite the developer knowing that it's unnecessary.

As an example where this is useful, see: https://github.com/shuklaayush/noir-bigint/blob/9a876290c9f75058dc53e74d25da5b3d79ada0d8/crates/biguint/src/utils.nr#L6-L21

In the line (ret as u112 >> BITS as u112) as u56 we're performing 2 casts (BITS as u112 gets handled at compile time so doesn't count) which automatically includes 2 truncations despite the fact that we know that these casts are valid and will not cause an overflow in the target type. If we had an unsafe cast then we could simplify this down to a single division rather than it having 3.

In order to do this we need to remove the truncation logic from a cast so that doing x as u8 would be represented in SSA as two instructions (a truncation and then a cast). Cast instructions would then not be represented in the ACIR at all but only exist to change the types of values in SSA.

Once this is done we can add some form of unsafe_cast function to the stdlib which would emit just a cast instruction without the truncation.