feat: reduce the number unconstrained boundary hops in bloblib

[TomAFrench]

I've taken a quick look through the bloblib to see if we can make any optimizations on the Noir code itself. This PR has a few changes from this:

• I've removed the usage of std::as_witness as all the usage was applied to the return values from unconstrained functions (and so will be witnesses anyway).
  • This shouldn't have affected performance but it also wasn't providing us with any benefits.
• Rather than calling many unconstrained functions from a constrained function I have created unconstrained helpers which will perform all calculations and then return any values which are necessary for applying constraints.
  • This minimises the amount of data which gets pushed from the ACVM to the Brillig VM and vice versa. This is important as we no longer need to read all of arguments from calldata and cast them to the correct types or arrange the results in memory to be read back into the main witness map by the ACVM.
  • This means that for the calculation of a single partial_sum, we would have to spin up and tear down the brillig VM 17(!) times whereas now it will be done just once.
• I've removed some cases were we'd use BigNum::evaluate_quadratic_expression in order to perform simple arithmetic and replaced them with the equivalent constrained operation. An example for this is calculating z_pow_d.

[github-actions]

Changes to circuit sizes

Generated at commit: 5a28753866b7b872006362aa263e2bc54c344388, compared to commit: 0332bb8893a880e3c05869b26ae45699ac3876c7

🧾 Summary (100% most significant diffs)

Program	ACIR opcodes (+/-)	%	Circuit size (+/-)	%
rollup_block_root	-50,688 ✅	-6.92%	0 ➖	0.00%

---

Full diff report 👇

Program	ACIR opcodes (+/-)	%	Circuit size (+/-)	%
rollup_block_root	682,326 (-50,688)	-6.92%	4,404,486 (0)	0.00%

[TomAFrench]

@MirandaWood looks like this is fixed now.