fix: don't read past position 0 when checking dollar escaping in dollarmath - #174

Fixes #173

Summary

dollarmath's is_escaped helper decrements its position index inside a while back_pos >= 0 loop, so for a dollar at position 0 the index becomes -1 and Python's negative indexing reads the last character of the inline source instead of finding no preceding character:

• Math at the very start of a paragraph was silently dropped whenever the inline source ended with a backslash — rendering $a$ is math\ produced literal $a$ instead of the math span (with \\ at the end the same happened).
• Backslash runs before a mid-source dollar could pick up trailing characters, flipping an odd count to even so an escaped dollar was treated as an opening delimiter.

The fix is the one-character bound: while back_pos > 0. Position 0 has no preceding character, so the count stays 0.

Testing

• test_inline_func_start_with_trailing_backslash — $a$ is math\ and $a$ is math\\ at position 0 must parse as math. Fails on un-fixed HEAD (rule returns False, no token).
• test_inline_func_escaped_dollar_mid_source_with_trailing_backslash — \$x$ y\ from position 1 must stay un-parsed. Fails on un-fixed HEAD (rule returns True and pushes a math token).
• Full suite: 636 passed, 3 skipped, 1 xfailed on Python 3.13, plus the 16 pre-existing data_regression fixture collection errors that also occur on the un-modified checkout (the pytest-regressions plugin is not installed in this environment).
• Re-verified the rendered output for all four shapes with the plugin active (math-at-0 kept, escaped mid-source dollar kept un-parsed).

Notes

The mid-source escaped-dollar path is normally masked in full parsing by markdown-it's own escape rule consuming the \$ pair first, so it is covered as a unit contract of the helper; the position-0 dropped-math case reproduces end-to-end through .render().