[feat] opcode offsets redesign

[Golovanov399]

This resolves INT-2355 to some extent.

While this doesn't eliminate the need for the offsets completely, this change presents an API for opcode mapping at the cost of need to define the new VmCoreAir trait function start_offset(&self) -> usize. After that one can use opcode_to_global_expr or expr_to_global_expr to map from local opcode to global in the air.

In most of the chips though the offset is no more stored and the default offset is used instead.

To convert from global to local, we can still use the VmOpcode::local_opcode_idx(offset) method.

Among other things:

• UsizeOpcode is renamed to LocalOpcode
• UsizeOpcode::default_offset() is now LocalOpcode::CLASS_OFFSET,
• UsizeOpcode::as_usize() is now LocalOpcode::local_usize() to avoid confusion,
• UsizeOpcode::with_default_offset() is now LocalOpcode::global_opcode(),
• All of the above, including the LocalOpcode trait itself, can be renamed easily.

[jonathanpwang]

LGTM except:

• some variables named class_offset should be start_offset
• the loadstore core chip's usage of LocalOpcode as a generic differs from how it's handled in e.g., ALU. is there a reason or is this some leftover?

[github-actions]

group	app.proof_time_ms	app.cycles	app.cells_used	leaf.proof_time_ms	leaf.cycles	leaf.cells_used
verify_fibair	(-31 [-1.3%]) 2,274	(-17268 [-3.4%]) 494,624	(-482045 [-2.5%]) 18,828,544	-	-	-
fibonacci_program	(-66 [-1.1%]) 6,088	1,500,137	51,505,102	-	-	-
regex_program	(-281 [-1.5%]) 18,685	4,190,904	165,028,173	-	-	-
ecrecover_program	(-16 [-0.6%]) 2,605	285,401	15,092,297	-	-	-

Commit: f18dcb6

Benchmark Workflow