Fixes to parser and explicitly print parsing error

[shubhamugare]

This PR includes two main changes:

1. There are some fixes in parsers when the generation starts with empty output in the case of NL to code tasks for general programming languages.

2. The parsing error and switch to unconstrained generation on an error is made more explicit. -- On any parsing error SynCode typically falls back to unconstrained generation (unless dev_mode flag is on). In the case of the grammar_mask mode, a parsing error can occur when the SynCode generation diverges from the grammar as SynCode does not always guarantee adherence to grammar.

Test:

python3 syncode/infer.py --mode grammar_mask --model meta-llama/Llama-2-7b-hf  --dataset humaneval --grammar python --max_new_tokens 400 --parser lr --parse_output_only False

results in pass@1 of ~12%