Preserve all streams when interleave inputs share an iterator

[fhgffy]

interleave can silently drop values from unrelated inputs when the same iterator occupies more than one input slot:

shared = iter([0, 1])
list(interleave([shared, shared, [10, 11]]))
# Current: [0, 1, 10]
# Expected: [0, 1, 10, 11]

If the shared iterator starts empty, the entire [10, 11] input is lost. The removal loop uses iterator identity to find the exhausted slot again, so another occurrence of that iterator ends the retained cycle too early.

Wrap each input occurrence in a distinct single-input itertools.chain. This preserves shared consumption while giving each slot its own identity. The existing cycle/removal algorithm and inner loop remain intact, including support for infinite inner and outer inputs.

The three regression tests cover adjacent and separated shared slots, empty and odd-length shared inputs, and infinite inputs. All three fail before the change and pass afterward.

Validation on Linux:

• CPython 3.12.14 and 3.13.5: full package tests and doctests each pass with 271 passed, 2 skipped
• pytest bench/ -q: 19 passed
• Repository CI's pycodestyle command and git diff --check: passed
• Separate round-robin oracle checks passed for 15,125 finite shared-iterator configurations, plus consumption/laziness and exception-propagation checks
• Other Python versions and platforms were not run locally

Performance tradeoff: the wrapper adds per-input overhead, and timings vary by workload. One local median-of-five sample for ten streams of 1,000 elements (number=200) took 0.0751 s before and 0.0930 s after, about 24% slower. Uneven-length and many-short-stream samples were close to baseline (ratios 0.963 and 0.995). This fixes data loss but does not claim performance neutrality.