prop-flow: stop reporting live props as caller-dead

[fxOne]

Summary

Two ways a prop that every call site passes could be reported caller-dead —
whose advice is "inline the default, remove the prop" — on live code, plus the
guard for the ways that are left. Nothing in the library API or the CLI
arguments moves and verdictOf keeps its signature; Site.kind gains a value.
Released as 3.1.0.

Changes

The two holes

• children is read off the JSX nesting rather than only the attributes. The
  value a call site writes between the tags was invisible, so
  <Panel><Body/></Panel> counted as a caller passing nothing and a children?
  nested at every call site came back caller-dead. Nesting now wins over an
  attribute of that name and over every spread, matching how JSX itself
  resolves it
• A lone <Panel>{slot}</Panel> is treated as a pass-through like any other
  value, so a forwarded children is traced to its source. Whitespace between
  the tags, a lone {/* comment */} and <Panel></Panel> reach nothing and
  stay omissions
• unwrapToFn takes an optional DeclarationResolver and follows a wrapper
  argument that only names the function: function CardComponent(…) {} plus
  export const Card = memo(CardComponent) was two components to the walk, so
  a pass-through climbing out of the body arrived at an empty usage list. The
  wrapped function may live in another file — memo(CardImpl) over an import
  resolves through the alias. A seen set guards the cycle that only
  half-written code produces
• The usage index keys on the component function a symbol resolves to instead
  of the symbol itself, so <Card/> and CardComponent land on one entry. The
  resolution is memoised per symbol — symbolId runs on every JSX tag and
  every callee in the Program
• The same gap made findComponents skip a file whose component is exported
  through such a wrapper, reporting no components at all for Card.tsx

The guard

• A caller-dead whose sites include a pass-through that contributed nothing
  is now reported manual. Both fixes above were one shape of the same
  failure: the subtree came back a silent 0/0/0, the omissions written
  elsewhere were all that was left, and the walk concluded nobody passes a prop
  that is passed on every render. Three reasons a subtree comes back empty —
  the component really is dead, it is called rather than rendered (already
  manual), or the walk filed its call sites under a different key. The first
  and the third are indistinguishable, and only the first may safely end in
  "delete this"
• Such a pass-through is reported under a new Site.kind, silent, and the
  hint names it, so the row says which line to go and check. The kind prints
  under every verdict as evidence; only caller-dead is downgraded by it, and
  the counts never move — a downgraded row still reads passes=0. A
  self-recursive pass-through is not a silence: the first visit counted that
  subtree, and the repeat is meant to add nothing
• Scoped to caller-dead on purpose. spread.tsx: Leaf.note keeps its
  justified verdict with a silent line in it, which is what settles the
  open question in prop-flow: downgrade caller-dead when a pass-through contributed nothing #939: the fixture's stance — "a dead component contributing
  nothing is the honest answer" — needed no renegotiation, because the
  downgrade never reaches a verdict nobody acts destructively on

Fixtures

• children.tsx, wrapped.tsx and wrapped-impl.tsx for the two holes;
  silent.tsx for the guard, carrying both a caller-dead that a silent
  pass-through pulls back and one that bottoms out directly and must not move
• Tests over the unwrap helper, the usage index, the analyzer and the hint

Related issues

Closes #939