feat(BREAKING): replace sqlformat with our own SQL formatter

[dsherret]

Replaces the sqlformat-rs wrapper with our own SQL formatter: a dialect-aware lexer, a tolerant parser and dprint-core's IR, following what was done in dprint-plugin-toml and dprint-plugin-markdown.

Closes #3
Closes #12

What it does

• Dialects: generic (the default), postgresql, mysql, sqlserver, sqlite, oracle, bigquery and snowflake, chosen with the dialect option. The dialect decides how comments, quotes, strings, variables and operators are read.
• Procedural code and stored procedures:
  • BEGIN/END blocks and exception handlers
  • IF/ELSIF, T-SQL IF/WHILE, loops, CASE statements and labels
  • declaration sections, cursors, Oracle packages and type bodies
  • PL/pgSQL $$ bodies, which are formatted as SQL too
  • MySQL DELIMITER, T-SQL GO batches, SQL*Plus / and its commands, psql meta-commands
• lineWidth: lists, conditions and brackets break only when they don't fit (Expose a config for maxInlineBlock / VALUES tuple wrapping #12).
• Set operations: EXCEPT and MINUS go on their own line like UNION and INTERSECT (set operations inconsistently formatted #3).
• Templates: dbt/Jinja tags ({{ }}, {% %}, {# #}) are kept as they are.
• Range formatting.
• Safety check: after formatting, the output is lexed again and compared with the input. The tokens must be the same, apart from keyword casing when casing changes it, and so must the line breaks that matter (between joined strings, and template tags in place). If anything else changed, the file is left as it is and an error is reported, so a formatter bug can't silently change code.

Breaking changes

• The output differs from sqlformat's.
• The global lineWidth (default 120) now affects SQL output. It was ignored before.
• linesBetweenQueries is now a minimum: existing blank lines between statements are kept.
• format_text returns a Result. Dialect and FormatError are #[non_exhaustive].
• The daily workflow that published a release for each new sqlformat version is removed, since there's nothing to follow anymore.

All existing options are still accepted, including the deprecated uppercase.

Testing

• Specs and tests: about 500 spec cases across every dialect and option, plus integration tests for errors, CRLF, ignore comments and performance.
• Every dialect: every spec input is also formatted in all 8 dialects at widths 40 and 120. Each result is checked for stability, and for only whitespace changing.
• Real-world SQL: about 3,100 files from the sqlfluff fixtures and the PostgreSQL, Oracle and SQL Server corpora, formatted in all 8 dialects at widths 30 and 120. No output changed on a second pass, nothing was refused because of a bug, and nothing crashed.
• Fuzzing: a multi-byte character is inserted at every position of every spec input, in every dialect, about 960k checks. No crashes and no unstable output.
• CI: now also runs clippy, cargo fmt --check and the tracing feature's tests.

Review

Four reviews (lexer, parser, generator, packaging) were done before opening this, and their findings are fixed in the follow-up commits:

• Snowflake stage paths, BigQuery dashed project names, # in MySQL variables, unicode whitespace
• statement splitting in T-SQL without semicolons
• SQL*Plus commands
• template tags around commas
• PostgreSQL joined strings
• exponential time on deeply nested CASEs and subqueries
• instability with a comment after a clause keyword

Known follow-ups, none of which change code:

• M1: when a single-action ALTER TABLE ... ADD COLUMN is too long, its type's brackets get broken (varchar(\n 255\n)). It should get the ALTER TABLE x\n ADD ... layout instead.
• M2: some lines are left past the width: long || chains, a trigger's EXECUTE FUNCTION, and GRANT ... TO a, b, c.
• M3: types are cased inconsistently with casing: upper: CAST(a AS text) but a::TEXT.
• M4: a few procedural constructs are laid out poorly. They are still verified and stable:
  • PL/pgSQL #variable_conflict
  • Oracle compound triggers and RETURN SELF AS RESULT
  • END LOOP label, and T-SQL END CONVERSATION
  • a comment between a label and its loop
  • a MERGE inside an exception handler that has a comment before its WHEN
• M5: in a T-SQL routine without BEGIN, linesBetweenQueries: 2 puts a blank line after AS.
• L1: {% endif %} at the end of a list loses its indentation.
• L2: /* dprint-ignore */ at the start of a line moves onto a line of its own.
• L3: newlines inside strings follow newLineKind, like the rest of the file.