FIX Added validation for TSNE.perplexity against data - #23471

For reference: https://www.jmlr.org/papers/volume9/vandermaaten08a/vandermaaten08a.pdf

The paper details that perplexity is used in defining a search criterion for the parameter $\sigma$. In particular, each choice of $\sigma_i$ induces a probability distribution $P_i$. Perplexity is defined as $\phi(P)=2^{H(P)}$ where $H$ is Shannon entropy. The user-defined perplexity parameter ($\Phi$) is used as a search criterion, meaning a binary search is performed on $\sigma$ until one such $\sigma_i$ is found that induces a distribution $P_i$ such that $\phi (P_i)=\Phi$. If we consider that the maximum-entropy distribution $P_*$ is the uniform distribution w/ $H(P_*)=\log_2(|X|)$ then the maximum perplexity is $\phi(P_*)=|X|$ and consequently setting $\Phi\geq |X|$ is problematic.

Edit: By the exact same reasoning, perplexity should overall be bound within $(1, |X|)$, however for convenience in maintaining the current tests that use one sample, we can validate in $[1, |X|)$ since it is unlikely anyone will reasonably choose $\Phi =1$. This validation can be done more rigorously in a follow-up PR doubling up to serve https://github.com//issues/23462

Edit: Linking back to the original stack overflow question referenced for this issue as an empirical study: https://stats.stackexchange.com/questions/332370/why-do-i-get-weird-results-when-using-high-perpexity-in-t-sne

Thanks @Micky774