Avoid retaining `TypeTable` in non-toplevel `StarlarkFunction`s

Once static type checking and toplevel evaluation of a .bzl file are complete, most of the data in `TypeTable` is no longer needed. Previously, every `StarlarkFunction` retained a reference to its file's `TypeTable`, keeping the entire `TypeTable` alive.

Instead, have `StarlarkFunction` store a shared array of function types for the current module (indexed by function ID) which we extract from the type table. (Strictly speaking, a `StarlarkFunction` needs to know only the type of itself and of the nested functions that it declares; but storing the entire array - as long as the array instance is shared by all functions in the module - results in simpler code, and wastes almost no memory in practice, as measured by benchmarks on large builds.) We still need to store `TypeTable` for `<toplevel>` and `<expr>` `StarlarkFunction`-s, but the toplevel `StarlarkFunction`-s are temporaries, and will get GC-ed after toplevel evaluation.

Also, take the opportunity to lazily initialize `TypeTable.typeConstructors` and `TypeTable.errors`.

Continuing to work towards minimizing the memory impact of enabling static type checking by default.

Working towards #27370

PiperOrigin-RevId: 1006090973
Change-Id: I556e0f600db1a4ac9aaaebae370107beafd83248