Build a name bind as a NameData, not as text

[sfc-gh-okalaci]

Stacked on #40.

NAMEOID was handled in the same case group as TEXTOID, VARCHAROID and
BPCHAROID, all of which construct a varlena. name is not a varlena — it is a
fixed-length NameData of NAMEDATALEN bytes with no length word. Building one as text
writes a varlena header into the first bytes of the value, and anything comparing it
against a real name reads that header as characters.

Nothing raises, which is what makes it expensive to find. The value simply never
matches:

pljs.execute('SELECT nspname FROM pg_namespace WHERE nspname = $1', ['pg_catalog'])
// 0 rows

pljs.execute('SELECT $1::name::text AS t', ['hello_name'])[0].t
// '8'

Catalog introspection by name — nspname, relname, typname, attname — is how
most of it is written, so the blast radius is wider than the round trip suggests: any
such query silently returns nothing.

NAMEOID now gets its own case and goes through namein(), which lays the value out
correctly and applies the truncation rule at NAMEDATALEN - 1 instead of raising. The
QuickJS string is released on the error path as well, so a value namein() rejects does
not leak it.

Test plan

sql/pg_name_bind.sql covers a catalog lookup by name, the round trip, the empty name,
truncation of an over-long value, and returning a name from a function.

It discriminates: with the conversion reverted, both catalog lookups return 0 instead
of 1, and the round trip returns 8 and \x10 instead of the names. Full suite green
on PostgreSQL 17; builds clean on 16, 18 and 19beta3.