Several places where the IEC grammar reduces a bare identifier through
multiple aliased or overlapping rules (type vs. function-block names,
derived type-name variants, access-path segments) are genuinely
ambiguous from syntax alone - which one applies is only decidable with
a symbol table - so those spots are collapsed onto the existing shared
`type_name`/`symbolic_variable` rules instead, deferring the
distinction to a later semantic pass.

Other conflicts are fixed more locally:
- signed_integer/real_literal fold their optional sign into the token
  regex so tree-sitter's longest-match lexing picks them over a bare
  unary-minus operator.
- bit_string_literal requires its type prefix, since an unprefixed
  integer was ambiguous with a plain integer literal.
- character_string (used where no STRING/WSTRING keyword disambiguates)
  is now a single unaliased rule instead of a single-vs-double-byte
  choice that shared the same delimiter and representation set.
- param_assignment's `NOT var_name => var` vs. unary_operator's
  `NOT expr` is a genuine local ambiguity (resolvable with more
  lookahead, not by restructuring), so it's handed to GLR via
  `%conflicts`.

Also drops fb_name_decl, made dead by the type_name consolidation.

Five GLR conflicts remained after the previous pass, all the same root
cause: a bare identifier reachable through more than one differently
labeled rule with nothing in the following tokens to tell them apart
(program_access_decl/access_declaration's type name, array_specification's
element type, data_source's resource/program/global-var prefix,
global_var_reference's own prefix/suffix, and _instance_specific_init's
fb_name vs variable_name). Each is resolved the same way the file already
resolves this pattern elsewhere: drop the redundant labeled alternative and
rely on the single already-reachable rule (type_name, symbolic_variable),
letting a symbol table (not this CFG) resolve which kind it actually is.
Dead rules left over from these fixes (non_generic_type_name and its
six-way identifier aliasing, program_output_reference,
function_block_type_name's derived-name half) are removed, and
standard_function_block_name's FIXME placeholder is filled in with the
fixed set of names from IEC 61131-3:2003 2.5.2.3 (SR/RS, R_TRIG/F_TRIG,
CTU/CTD/CTUD variants, TP/TON/TOF), with a new call site for it.

Also adds two pieces of parser infrastructure that were missing entirely:
`%extras` for `(* ... *)` and `//` comments (real ST code is full of them),
and a fix for the generated parser's start rule, which had no explicit
axiom and so defaulted to whichever rule happened to land first after
%include flattening - silently `elementary_type_name` rather than
`library_element_declaration`.

Replaces tests/literals.txt, which had gone stale (written against an
earlier structure of the grammar and no longer matching current rule
names/shapes) and had never actually been exercised, since ts-bnf-tool
generation was failing on the conflicts fixed in the previous commit.

tests/community_samples.txt has six small, complete, hand-written ST
declarations pulled from real MIT-licensed open-source projects (not
AI-generated) - WengerAG/structured-text-utilities and tkucic/UniTest -
covering a spread of rules (enum/struct TYPE, functions with nested
calls, VAR_TEMP, IF/ELSE, a function block with FOR/array/struct-field
access). tests/SOURCES.md has the license/attribution details.

Only one of the six currently parses clean; the rest hit real grammar
gaps the corpus surfaced (mandatory `;` after already-block-terminated
constructs, enumerated_value's mandatory qualifier, STRING[n] as a
general type name, FUNCTION's single non-repeatable, VAR_TEMP-less
declarations slot). Documented in tests/FINDINGS.md with examples, root
causes, and recommended fixes, held for review before changing the
grammar further.