Decide how to handle non-deterministic t-SNE algorithms

[csadorf]

PR #8123 closed #2980 by making fixed-seed method="fft" runs byte-deterministic. It did not make method="exact" or method="barnes_hut" deterministic; both can still produce different results with the same random_state because their GPU implementations use nondeterministic parallel operations.

We should decide how to handle the remaining algorithms:

• Characterize reproducibility, performance, embedding quality, and distinct use cases for exact and barnes_hut.
• Estimate the cost and performance impact of making each implementation deterministic.
• Review whether maintaining all three algorithms is still justified now that FFT is the default and the only deterministic implementation.
• Consider deprecating and eventually removing exact and/or barnes_hut as an alternative to adding deterministic execution paths.
• If an algorithm remains supported, document its reproducibility guarantees and add appropriate coverage.
• If deprecation is selected, define the migration guidance, warning period, and removal release.

Follow-up to #2980 and #8123.