This PR does a few things:

1. Fixes peephole optimizations not being applied to opcodes created by peephole optimizations. E.g.:
   PushNull -> Assign -> Pop was being optimized to PushNull -> AssignPop, but the optimizer offset wasn't rolled back far enough to consider optimizing PushNull -> AssignPop to NullRef. It would only be rolled back to the new AssignPop opcode. This meant that the NullRef opcode was never used, which is now fixed.

2. Fixes a few opcodes not being implemented in ProcDecoder. Because a few opcodes were specific to optimizations that were never applied, they never triggered any errors in the disassembler. DreamProcOpcode.NullRef and DreamProcOpcode.PushNOfStringFloats can now be decoded.

3. Fixes PushNOfStringFloats only being applied to pairs of StringFloats. Because of how the peephole optimizer works, it would (now that it actually runs) try to run the PushNOfStringFloats optimization before every PushString and PushFloat in the sequence had been converted to StringFloats. The PushNOfStringFloats optimization has been folded into the StringFloat optimization, and it'll just use the relevant opcode depending on how many pairs of PushString and PushFloat opcodes there are. This has the added benefit of reducing the amount of optimizations that need to be applied.

We really need to implement a system for unit testing the peephole optimizer on input bytecode, but that is out of scope for this PR.

I made sure that TG doesn't explode, the NullRef opcode is now used, and that the PushNOfStringFormat optimization actually works correctly & disassembles with sane formatting:

	PushFloat 0 
	PushNOfStringFloats "y" 0 "size" 1 "repeat" 2 "radius" 0 "falloff" 1 "flags" 0
	CreateAssociativeList 7