Reject a byte position the lexer cannot start on

[soutaro]

rbs_lexer_new reaches start_pos by stepping one character at a time from the beginning of the buffer, so the only positions it can start lexing from are the first bytes of characters. A byte_range can hand it two other kinds of position:

Inside a character. The step walks over start_pos and lexing quietly begins at the next character boundary. Nothing downstream can tell: line, column and char_pos all stay consistent — they just describe a position the caller did not ask for. Today the shift often surfaces as a syntax error by accident, because a non-ASCII character cannot open a token; but an ASCII letter right after a multibyte character parses cleanly in the wrong place, and #3082 gives non-ASCII characters tokens of their own, which takes even the accident away.

Past the end of the buffer. There is nothing left to step over, so the walk never finishes: rbs_skip stops moving at EOF while the loop waits for a byte position it will never reach.

RBS::Parser.parse_type("Integer", byte_range: 20...30) # hangs on master

This PR makes the walk report an unreachable start_pos: rbs_lexer_new and rbs_parser_new return NULL, and the C extension turns that into an ArgumentError alongside the existing reversed/negative range validation, since both are the caller's mistake rather than anything about the source text.

RBS::Parser.parse_type('"🐕🐈"', byte_range: 2...)
# => ArgumentError: position range starts inside a character: 2...10

RBS::Parser.parse_type("Integer", byte_range: 20...30)
# => ArgumentError: position range starts past the end of the buffer: 20...30, buffer is 7 bytes

end_pos keeps accepting any value: clamping with a large number instead of measuring the buffer is ordinary, and the lexer stops at the end of input on its own.

Note: this overlaps with #3082 in test/rbs/type_parsing_test.rb — both rewrite test_parse__byte_range_incorrect — so whichever lands second needs a trivial rebase.

🤖 Generated with Claude Code