fix(fmt): preserve syntax across formatting round trips - #1523

Master keeps attributes on for loops now; this adds cases it does not
cover: several attributes stacked on one loop, attributes on a nested
loop, and irregular spacing in the loop header.

…tributes

Inner attributes of a nested module (`#![...]`) were printed from a
separate path that dropped the blank line between them and the first
item. Render them like other attribute lists instead: keep the line
breaks that follow them, and print `#![` for inner attributes in
`NormalAttr`. This also lets the next change keep comments placed among
inner attributes.

An attribute list was formatted from its attributes only, so a plain
comment between doc comments and attributes was deleted (as in
`std/src/abi/sol.fe`).

A comment right after the attribute list belongs to the item node, so it
sent the whole item through the comment-preserving fallback, which
printed it with a leading space. Such comments are now kept with the
attribute list, one per line, and the item is formatted as usual.

The one-element tuple rule also applied to tuple variants, so `Some(T)`
and the pattern `Some(x)` were written as `Some(T,)` and `Some(x,)`.
Only a one-element tuple type or pattern needs that comma.

`pub(ingot)` and `pub(super)` still parse after being widened to `pub`,
so only a snapshot catches that regression. Cover functions, structs,
fields, impl functions, consts, enums, traits, type aliases, modules,
`use` and extern functions, with and without comments inside the item.

The corpus test only checked that formatted files parse and are stable,
which passes when the formatter deletes an attribute, a comment or a
visibility restriction. Also compare the tokens of each file before and
after formatting, ignoring whitespace, separator commas and the braces
added around `return` match arm bodies.

The visibility, for loop attribute and inner attribute fixes already ship
from master with their own entry, so only mention what this branch adds.