Stop the lexer reading past the end of a byte_range

[ksss]

rbs_next_char ends the input when byte_pos == end_pos. rbs_skip advances by a whole character, so a multibyte character starting before end_pos and ending after it steps over the boundary — equality never holds again, and the lexer reads to the end of the string.

The realistic way to land inside a character is to pass a character offset where a byte offset is expected — the mistake #2945 fixed in parse_inline_*_annotation. "日本語" is 5 characters but 11 bytes, so offset 5 falls inside 本, which spans bytes 4...7:

RBS::Parser.parse_type('"日本語" | Integer', byte_range: 0...5)
#=> RBS::Types::Union spanning the whole input, rather than an error

require_eof: true does not catch it, because by then the lexer really is at EOF.

It takes a character straddling the boundary, so ASCII-only input never hits it. On master today it also needs that character to sit somewhere non-ASCII is already accepted — a string literal or a comment — since a non-ASCII byte in an identifier stops at an ErrorToken first. #3028 lets identifiers hold non-ASCII, which widens the set of inputs that reach it, but the bug is independent of that PR.

Fix

Compare with >= so stepping over the boundary still ends the input.

That leaves the straddling character itself read whole, one character past the range. On master that is not observable — every case I could construct still ends in a parse error — so nothing more is done here. It becomes observable once an identifier can hold non-ASCII, since an identifier is a complete type on its own; #3028 tightens the bound there, where a test can cover it.

Test plan

• rake test — 968 tests, 0 failures. Two errors also present on master, unrelated.
• New test_parse__byte_range_ending_mid_character, confirmed to fail with the == comparison restored. Covers the character-offset mistake above, the byte offset the caller meant, and an end_pos past the end of the buffer.