_dialect.py Re-Work, modifier Argument for LLMQA, Documentation updates

[parkervg]

• Re-worked _dialect.py to better handle different SQL dialects
  • Previously, we forced sqlglot to interpret all queries as SQLite, and relied on the serendipitous interpretation of ingredients as STRUCT nodes to do our processing
  • Now, we have the get_dialect() function to route queries to the correct sqlglot dialect, and interpret all aliased ingredients (i.e. {{A()}}, {{B()}}, etc.) as a TokenType.FUNCTION via INGREDIENT_TOKEN_TYPE_MAPPING
• Added support for modifier argument
  • LLMQA now works in a context where a list datatype is required, such as IN or VALUES clauses
  • To unify handling of data types across ingredients, the DataType object in _constants.py was created
• Added AzurePhiModel
  • Useful for testing constrained decoding with guidance via a stronger remote model, vs. very very tiny local cpu-bound models
  • https://github.com/guidance-ai/guidance?tab=readme-ov-file#azure-ai
• Added blendsql-by-example.ipynb

Before:

Task	Average Runtime	# Unique Queries
financials	0.0440116	7
rugby	1.26541	4
national_parks	3.77385	5
1966_nba_draft	4.02796	2

After:

Task	Average Runtime	# Unique Queries
financials	0.0415093	7
rugby	1.18449	4
national_parks	3.00202	5
1966_nba_draft	4.00881	2