[MIR] Speedup parsing of function with large number of basic blocks

Patch eliminates string length calculation when lexing a token. Speedup can be up to
1000x.

Differential revision: https://reviews.llvm.org/D83389