make `TaskInput::is_resolved` inlinable everywhere (#94213)
## What
Removes the `ArgMeta::is_resolved` function-pointer field and the 592
monomorphized `closure#3` trampolines it generated. The
`TaskInput::is_resolved` check that runs on the synchronous fast path of
every turbo-tasks function call now happens at the macro-generated
callsite on the concrete tuple type, before erasure into `&dyn
StackDynTaskInputs`. For trait calls with `filter_owned`, the check is
fused into the filter functor so the filter and the resolved-check
monomorphize together.
## Why
The `is_resolved` method is almost always trivial after optimization but
putting it behind an opaque function pointer means we always pay a small
code size and perf cost. Inlining actually saves bothr `Vc` field.
Bonus rename: `OwnedStackDynTaskInputs` → `BoxedDynTaskInputs`. The
\"Stack\" was the trait, not the storage — the inner value is a `Box<dyn
DynTaskInputs>` on the heap.
## Hot-path microcheck
The synchronous wrapper for a representative 6-input turbo function
(`directory_tree_to_loader_tree`) now performs the entire `is_resolved`
check as ~15 register-only ARM instructions inlined into the caller —
three `cmp/csel/tbnz` chains testing the discriminant byte of each
`Vc<...>` input. The four trivially-resolved inputs (`FileSystemPath`,
`RcStr`, `AppPage`, `AppPath`) contribute zero instructions because
their `is_resolved` constant-folded to `true`. No indirect call, no
downcast trampoline.
## Code-size impact
Release `next-swc.darwin-arm64.node`, macOS aarch64, vs the `temporary`
branch this PR is stacked on:
| Metric | prior (temporary) | this PR | Δ bytes | Δ % |
| --- | ---: | ---: | ---: | ---: |
| Raw `.node` | 124,217,280 | 123,845,712 | −371,568 | −0.30% |
| Stripped (`strip -x`) | 84,013,136 | 83,880,384 | −132,752 | −0.16% |
| Stripped + gzip -9 | 28,898,920 | 28,887,624 | −11,296 | −0.04% |
<!-- NEXT_JS_LLM_PR -->
