diff: don't panic on an oversized context count

[MsfPablo]

Fixes #245.

Both panics in the issue come from the context count being used unchecked:

• -u99999999999999999999 and friends reach numvalue.as_str().parse::<usize>().unwrap() in params.rs, and PosOverflow aborts the process.
• A count that does fit in usize but is large enough that N * 16 > isize::MAX gets past the parser and aborts in VecDeque::with_capacity instead, which is the second case @leeewee added in the comments.

The parse now saturates, and make_diff clamps context_size to the combined length of the two inputs. Neither file can have more lines than it has bytes, so a request above that already meant "the whole file" — the clamp only removes values that were never reachable, and it keeps both the preallocation and the context_size + 1 arithmetic below it in range. That covers the unified and context paths in one place rather than guarding each site.

All six affected spellings now produce the same output as a plain -u/-c run and exit 1.

Tests: a unit test in params.rs for the saturating parse, and an integration test running each spelling end to end.

One thing I left alone: the separate-argument forms (-C 99999999999999999999) already return invalid context length rather than panicking. GNU accepts those, so the behaviour still differs, but since it is not a crash I did not want to fold a compatibility change into this fix. Happy to do it separately if you want them aligned.

[github-actions]

GNU diffutils testsuite comparison:

Test results comparison:
  Current:   TOTAL: 33 / PASSED: 0 / FAILED: 33 / SKIPPED: 0
  Reference: TOTAL: 33 / PASSED: 8 / FAILED: 21 / SKIPPED: 4

Changes from main branch:
  TOTAL: +0
  PASSED: -8
  FAILED: +12

New test failures (12):
  - basic
  - bignum
  - brief-vs-stat-zero-kernel-lies
  - bug-64316
  - cmp
  - diff3
  - help-version
  - large-subopt
  - strcoll-0-names
  - strip-trailing-cr
  - timezone
  - y2038-vs-32bit