Redisign dialect check

[jongleb]

Description

This PR eliminates indirection in the parser that previously relied on Parser_state.Dialect_feature to remember dialect-specific tokens for deferred processing. Instead, the frontend AST is now enriched to carry all necessary information directly.

Details

Previously, the parser used a mutable state module Parser_state.Dialect_feature to track dialect-specific features during parsing:
When encountering features like STRAIGHT_JOIN, REPLACE INTO, ON DUPLICATE KEY, collations, unsigned types, etc., the parser would call functions like set_straight_join, set_collation, set_unsigned_types to store feature markers in a global ref
These markers were later processed separately from the AST
This created an indirect flow where AST construction and dialect feature collection were decoupled, making the code harder to reason about and maintain.

Some type information is only relevant for dialect checking and should not propagate into the unification/resolution phase (res_expr) or code generation stage.
For example, UInt32 needs to be tracked for dialect validation, but at runtime it's always represented as int64 in OCaml. We don't want this distinction to leak into type inference or generated code.
To address this, we introduce Source_type.t:

 type kind = Infer of Type.kind | UInt32  type t = { t : kind; nullability : Type.nullability; }   

The naming Source_type (open to discussion) emphasizes that these types exist only at the "source" level — in the frontend AST — and are converted to regular Type.t before entering unification. This keeps dialect-specific details contained in the parsing/validation layer without polluting the core type system.