load_csv: skip blank lines, surface line number on row-length and key-column errors

[HrachShah]

Fixes #29.

A trailing blank line in a CSV file (the kind GitHub and most editors emit by
default) crashed the diff with a bare KeyError on the key column, far from the
actual problem. csv.reader yields an empty list for a fully-blank line, and
dict(zip(headings, [])) returned {}, so the next line raised KeyError('a')
inside the keyfn lambda. Issue #29.

While reading the file, track the 1-based source line number alongside each row
and use it to surface clear, line-numbered ValueError messages for the other
two error paths that previously leaked as tracebacks:

• rows with fewer fields than the header (the previous list-comprehension
  silently accepted them and produced dicts with missing keys, then crashed
  inside the keyfn lambda or in compare()'s next(iter(...)));
• a --key column that isn't in the header (was also a KeyError, this time
  on r[key] inside keyfn).

Trailing and interior blank lines (csv.reader yields [] for them) are now
silently skipped, matching the POSIX text-file convention and the behaviour of
most other CSV tools. Rows with more fields than the header are still accepted,
since the trailing-comma pattern is a real-world way to express an empty last
column; the FIVE test fixture exercises that.

Added five tests in tests/test_csv_diff.py:

• test_trailing_blank_line_ignored (issue BUG: new line at end of file causes crash #29 reproducer)
• test_interior_blank_line_ignored
• test_trailing_blank_line_with_no_key
• test_mismatched_row_length_raises_clear_error
• test_missing_key_column_raises_clear_error

All 29 tests pass (5 new + 24 existing).