Accept only SQL-standard type names in as_sqlr_type()

[nbenn]

Summary

• Type strings must now be one of the SQL-standard spellings from the issue's table, in any case and spacing. Anything else, engine aliases and a bare float included, is an error pointing to ?as_sqlr_type and sqlr_other(), instead of a silent sqlr_other().
• The new parser also fixes three misreadings: the old one dropped with time zone, read binary(n) as variable-length and left char unsized. A parsed type now equals what its constructor builds, with raw left for reflection to set, as ?sqlr_type documents.
• Types gain a format() method that runs the same table in reverse and falls back to the constructor call, such as sqlr_integer_type(bytes = 1L). A type difference in sqlr_diff() now reads varchar(255) vs text, and equality compares constructor calls, replacing type_key().
• The package now calls S7::methods_register() in .onLoad(). Without it an installed sqlr never registers S7 methods on base generics, so format() would only have worked under load_all(), and the existing print() method for sqlr_sql has been ignored all along. The format tests guard this under R CMD check, which runs them against the installed package.
• No dialect implements sqlr_parse_type(), so it is dropped, along with the utils import only the old parser used. Neither dialect has called as_sqlr_type() since Reflect column types with this dialect's own mapping sqlr.sqlite#2, so the stricter parsing reaches authored strings only.

Spellings that parsed before and are now refused:

Refused	Write instead
int2, int4, int8	smallint, integer, bigint
float4, float8, double, bare float	real, double precision or float(p)
bpchar	char(n)
bytea	blob
bool	boolean
timetz, timestamptz	time with time zone, timestamp with time zone
jsonb	sqlr_json(binary = TRUE)
varchar without a length	varchar(n), or text if unbounded
Unmodelled types, such as geometry	sqlr_other("geometry")

Fixes #9