test: measure the multi-mode font preload instead of arguing it

[awdr74100]

Follow-up to #168, which shipped the text-style font preload with a hole in its evidence: the
account it was built on caps a variable collection at one mode (in addMode: Limited to 1 modes only), so the multi-mode reasoning was argued rather than observed. The PR said as much. A
multi-mode account is available now, so it was measured.

What the measurement says

tried on a two-mode collection	result
fontFamily bound to a variable with Roboto / Lato per mode	binds, style resolves to Roboto
same, but the non-default mode names a font nobody has installed	still binds
fontFamily and fontStyle both bound, both multi-mode	binds, composes to Roboto Bold

The middle row is the useful one: only the face the style actually resolves to has to be
loadable. That makes swallowing a failed preload load-bearing rather than defensive — letting
that failure through would reject a binding Figma accepts. #168's comment offered that as a
plausible reason ("a face this guesses at may simply not exist"); it is now a measured one.

The third row confirms what the family → style chain exists for: the two bindings compose into a
single face, so the intermediate face has to be loaded for the family swap to land before the
style swap runs.

What changed

• The comment says what was measured instead of what was reasoned.
• Three tests pin it. Worth noting how the first attempt failed: the multi-mode test asserted a
  family that happened to sit in the first mode, so a mutation narrowing the preload to a
  single mode passed it. Rewritten with a second-mode family that must also be loaded, the
  mutation is caught.
• The faces are keyed while passing through: with a family binding and no style binding, the
  intermediate and final face are identical, and asking twice was pointless even against a cached
  loadFontAsync.

No behaviour change — the implementation was already right. This replaces an argument with
evidence, and closes the gap the tests had.