[CIR] Add custom constructor declaration to CIR lowering TableGen (#176664)

While working on the lowering of `BlockAddressOp`, I ran into a
limitation of the current `CIRLoweringEmitter` infrastructure: some CIR
ops require shared mutable state across multiple lowering patterns, but
the generated patterns always use a fixed default constructor and cannot
accept additional context.

**Examples:**

* `BlockAddressOp` needs to coordinate with the lowering of its
referenced `LabelOp`, requiring shared bookkeeping (e.g.
`LLVMBlockAddressInfo`) to track lowered labels and resolve unresolved
block addresses once the corresponding `BlockTagOp` is available.
* This is not unique to block addresses: for example,
`CIRToLLVMAllocaOpLowering` will also need access to shared state such
as `stringGlobalsMap`, `argStringGlobalsMap`, and `argsVarMap`.

**Possible solution:**

* **Disable generated lowering**: mark the op with `hasLLVMLowering =
false` and manually define and register the lowering patterns. This
works, but feels misleading since the operation does have an LLVM
lowering it just cannot be generated automatically. This ambiguity also
suggests that `hasLLVMLowering` might be better named `genLLVMLowering`.
* **Custom constructors (this PR)**: extend the TableGen emitter to
allow defining a single custom constructor for the generated LLVM
lowering pattern, similar to MLIR’s `OpBuilder`. When a custom
constructor is provided, the default one is not generated and the
pattern must be manually registered in the RewritePatternSet.