fix out-of-bounds read in antlr3UTF16LA backward surrogate peek

[zayeem06]

antlr3UTF16LA handles negative (lookbehind) LA by scanning backwards, and when the previous unit is a low surrogate it peeks at *(nextChar-1) to see whether a matching high surrogate precedes it. If that backward scan reaches the start of the buffer the peek reads input->data[-1], two bytes before the allocation; this happens with UTF-16 input whose first code unit is a lone low surrogate followed by a negative LA. The forward branch already bounds the equivalent surrogate peek, so I added the matching lower-bound check before the backward peek. I confirmed the read under AddressSanitizer (heap-buffer-overflow of size 2, two bytes before the buffer) and verified it is gone with the guard in place.