[CIR] Lower aligned operator new in CIRGen (and matching cleanup-during-new) (#197094)

Two related `errorNYI`s in CIRGen for C++17 aligned `operator new` /
`delete`:

- `emitCXXNewExpr` `errorNYI`'d on `"emitCXXNewExpr: pass alignment"`
when the allocator is the `align_val_t`-overloaded form (called whenever
you `new` an over-aligned type).
- The cleanup-during-new path (the `CallDeleteDuringNew` cleanup that
fires if a ctor throws after `operator new` returns) `errorNYI`'d on the
matching `"CallDeleteDuringNew: aligned allocation"`.

The non-cleanup delete path was already wired up; this just plumbs the
matching new side.

Both fixes are direct ports of `clang/lib/CodeGen/CGExprCXX.cpp`. In
`emitCXXNewExpr` we look up the alignment parameter type from the
allocator prototype (or fall back to `size_t` for the variadic corner
case classic also handles), build a constant of that type, and append it
after the size argument. The cleanup uses `getSizeType()` for the
alignment param type per the existing `// FIXME` comment that's also in
classic.

Hit this trying to build `706.stockfish_r` from SPEC CPU 2026 with
`-fclangir` — stockfish allocates over-aligned NNUE state via plain
`new`. The minimum repro:

```c++
struct alignas(64) Big { char data[64]; };
void f() { Big *p = new Big; delete p; }
```

The new test cross-checks against OGCG so any drift between the two
backends shows up — both emit the same `_ZnwmSt11align_val_t(64, 64)`
and `_ZdlPvmSt11align_val_t(p, 64, 64)` calls.

A few honest caveats worth flagging:

- The `CallDeleteDuringNew` fix isn't directly exercised by the test.
Triggering it needs `-fcxx-exceptions` plus a throwing ctor, and CIR's
EH path has its own NYIs that would muddy the test. The cleanup is a 1:1
mirror of the regular-delete-side fix that already landed and uses the
same `getSizeType()` fallback as classic.
- The variadic `operator new(size_t, ...)` corner case is also untested.
The `if (allocatorType->getNumParams() > paramsToSkip)` branch covers
the common case; the else branch falls back to `sizeType` (mirroring
classic, which guards the equivalent with
`assert(allocator->isVariadic())` and likewise has no test).
- The cleanup path uses `getSizeType()` for the alignment param type
rather than the actual `align_val_t` enum, mirroring the existing `//
FIXME` in classic. ABI-equivalent — `align_val_t`'s underlying type is
`size_t` — but the IR type label differs.

Both CIR test suites pass cleanly (`check-clang-cir-codegen` 325/326 + 1
unsupported, `check-clang-cir` 718/729 + 1 XFAIL + 10 unsupported).