nll should respect lifetime annotations from patterns

[nikomatsakis]

The following test passes NLL, but ought not to:

#![feature(nll)]

struct Foo<'a>(&'a u32);

fn main() {
    let y = 22;
    let foo = Foo(&y);
    let Foo::<'static>(z) = foo;
}

As does this one:

#![feature(nll)]

struct Foo<'a> { field: &'a u32 }

fn main() {
    let y = 22;
    let foo = Foo { field: &y };
    let Foo::<'static> { field: z } = foo;
}

Broken out from #47184

[nikomatsakis]

This ought to be able to re-use the existing pattern ascription nodes in the HAIR. I imagine we will store the substs into the user_substs field and then read them out when constructing HAIR, similar to what we are doing for ADT expressions:

rust/src/librustc_mir/hair/cx/expr.rs

Lines 293 to 300 in 31789a6

	let user_ty = cx.tables().user_substs(fun.hir_id)
	.map(|user_substs| {
	user_substs.unchecked_map(|user_substs| {
	// Here, we just pair an `AdtDef` with the
	// `user_substs`, so no new types etc are introduced.
	cx.tcx().mk_adt(adt_def, user_substs)
	})
	});

We can then generate a AscribeUserType pattern node:

rust/src/librustc_mir/hair/pattern/mod.rs

Lines 69 to 72 in 31789a6

	AscribeUserType {
	user_ty: CanonicalTy<'tcx>,
	subpattern: Pattern<'tcx>,
	},