Support arbitrary censoring via set_subtensor

[ricardoV94]

As the TODO comment describes, those types of graphs can be equivalent to clipping for the purpose of logprobs

aeppl/aeppl/truncation.py

Lines 25 to 27 in 2581f56

	@node_rewriter(tracks=[Elemwise])
	def find_censored_rvs(fgraph: FunctionGraph, node: Node) -> Optional[List[CensoredRV]]:
	# TODO: Canonicalize x[x>ub] = ub -> clip(x, x, ub)

More generally, we could also allow arbitrary censoring binning encoding of the form x[(x > lower) & (x < upper)] = encoding, but that would require a different logprob definition.