Interactive mode: an answer applies to one occurrence, not the whole run

[Eljees]

Fixes #62.

The root cause is that ask_for_word_fix answers a question about one occurrence by mutating the shared Misspelling object, so a single answer leaks into every later occurrence of the word in the run. Observed on current master (all transcripts scripted, answers fed on stdin):

• -w -i 1, the same word in two files, answers n, y: the second file never asks — the rejection turned the word off globally, and the queued y is never consumed.
• -w -i 3, reject a word, then meet it in the next file: the prompt degrades from (Y/n) to Choose an option (blank for none): 0) ... — the "handled as a list of available fixes" behaviour from the issue, still alive.
• -w -i 2, a word with a single candidate, answer blank ("blank for none"): the file is changed anyway — FIXED: c.txt.

Three changes:

1. ask_for_word_fix no longer mutates misspelling; it returns the answer for this occurrence.
2. The caller remembers answers per word per file (asked_for becomes a dict and moves out of the per-line scope, where its reset meant "asked once per file" only held because of the mutation). Accepting and rejecting now behave symmetrically: each word is asked once per file, and the answer covers the rest of that file.
3. The option-list branch now triggers only for words with more than one candidate, which is what --help says level 2 is for ("ask user to choose one fix when more than one is available"). Single-candidate words at -i 2 are written without a prompt, as documented; the blank-accepts-anyway path is gone.

These are the first tests for interactive mode in the suite; they feed answers through stdin. On the unpatched tree the three bug tests fail exactly on the behaviours above, the per-file control passes; with the fix all four pass. Full test_basic.py: 86 passed, 3 failed — the same three failures (chardet import) as on a clean tree. ruff check, ruff format --check clean; mypy reports the same two pre-existing errors as on a clean tree.

AI-assisted (LLM used for drafting); the runs above are mine.