Expose the parsed expression tree from Program (for explain/trace tooling, linting and sub-expression evaluation)

先检查 Python Program wrapper 和 cel::Program::expression()，然后将其 IdedExpr 和 Expr 变体与提议的可序列化为 JSON 的树进行比较。当 Program.ast() 暴露字面量、标识符、select、调用、列表、映射、struct 和展开后的 comprehension，并提供与 cel-rust 版本关联的临时文档时，即视为完成。

Motivation

Program.references() answers "which names does this expression use", but several things people ask for need the structure of the expression, not just its leaves:

• #30 wants to know which element or clause made an exists/&& chain true or false. cel-rust has no evaluation hooks, so a wrapper-side explain() would have to walk the tree and re-evaluate sub-expressions.
• Policy linting beyond name lists: "no matches() on untrusted input", "no more than N nested comprehensions", "field access only under request.".
• Pretty-printing / normalising expressions stored by users.

What upstream gives us

cel::Program::expression() returns the parsed Expression (IdedExpr: an id plus an Expr enum covering literals, identifiers, selects, calls, lists, maps, structs and the expanded comprehensions). No protobuf AST, but the shape is close to cel.expr.Expr and would be straightforward to mirror as nested dicts or small dataclasses.

Proposal

Program.ast() returning a plain, JSON-serialisable tree, e.g. {"kind": "call", "function": "_&&_", "args": [...]}, with comprehension macros shown in their expanded form (that is what cel-rust holds after parsing; the original macro call is not retained). Ship it as provisional, tied to the cel-rust version, and note that the node shapes may change when cel-rust changes its AST.

A Python-side cel.explain(program, context) built on it could then evaluate each boolean sub-expression and report the first failing clause, which covers the exists/all case in #30 without upstream changes, at the cost of re-evaluation (custom functions run again).