Collection of performance improvements

A list of improvements that may be worth investigating once we've got a proper benchmarking framework.
This list is to make sure we don't forget, nothing here has been investigated thoroughly.
Feel free to directly edit this message if you see something worthy.

• Use bytemuk for byte encoding/decoding (discussion).
• Save one byte in the CreateOrder transaction (#20).
• Drop the namespacing seed (discussion).
• Can we do better than reading all instructions in Begin/FinalizeSettle? Discussion.
• Should we revisit how the index is carried in Begin/FinalizeSettle? Since we scan anyway, should we only have a single index? The difference is expected to be small anyway. See discussion and that PR's description.
• Sanity check: is it true that allocating a constant for a hardcoded constant array doesn't decrease the CU cost? See discussion.
• Store bumps in the PDAs rather than passing them along in the transaction data (discussion).
• Large state PDA data has the side effect of unnecessary extra CU. This is true for both extra solvers and extra authorities. See the related discussion for some initial investigation.
• Better writing of accounts (vibed example: 21f11b948e9930a07f28e0f1c19e6f09473dccae)

Start by reviewing the linked discussions and related pull requests for one unchecked item, especially the notes about CreateOrder, Begin, and FinalizeSettle. A proper benchmarking framework is not identified here, so first establish how the chosen change will be measured. Done means one narrowly scoped improvement is investigated, benchmarked, and documented with evidence.