Cranelift: support different code models (i.e., relocation strategies and displacement limits)

Right now, we do not have a notion of "code model", i.e., a set of definitions for ways in which external symbols are referenced. The code model affects which types of relocations we emit, and which address-generation strategies we use: for example, going through a global offset table (GOT) vs. embedding addresses directly, or using PC-relative addressing modes vs. not, or using instructions with differently-sized displacement limits. The closest concept that we have seems to be the `colocated` flag on external function references.

Ideally, we would like to add support for this as a global compiler flag so that it can influence codegen.

This came up in #1570; non-x86 architectures (such as AArch64) will need it more urgently as the displacement limits are often lower (e.g., on AArch64, +/- 128MB for calls without going through some sort of indirection or veneer).

Start by reviewing Cranelift's existing `colocated` flag for external function references and the discussion in issue #1570. Define how a global code-model setting should control relocation strategies, address-generation choices, and displacement limits across supported architectures. Done means the compiler has a coherent code-model notion that influences code generation as described.

Written by the indexing model from the issue text.