Consider moving "Not in HLO" ops to CHLO

[burmako]

In https://discourse.llvm.org/t/rfc-proposal-for-a-high-level-ml-dialect-in-mlir/64249/46, we did a survey of MHLO ops (almost of that applies to StableHLO, since we've bootstrapped it from MHLO) and identified the following ops which are decomposable to other MHLO ops: broadcast, create_token, cross-replica-sum, dot, einsum, torch_index_select, unary_einsum. Let's move these ops to CHLO.

Note: In MHLO, broadcast and dot are referring to less general versions of broadcast_in_dim and dot_general. In HLO, broadcast and dot are used to refer to the more general versions, so we have a bit of confusing naming here which this move will also help to solve.

[subhankarshah]

BroadcastOp can be lowered to stablehlo.broadcast_in_dim for static shaped tensors and stablehlo.dynamic_broadcast_in_dim for ranked dynamic shaped tensors, however, It is not possible to decompose broadcastOp into any other stablehlo op for unranked tensors. We wait for the release of dynamism RFC to decide upon the migration of broadcastOp.

[burmako]

In retrospect, this ticket was a bit ahead of its time:

1. The fact that we haven't established a baseline specification prevented us from effectively planning the move. For example, as discussed above, it turned out that we cannot move broadcast to CHLO, because unranked broadcasts cannot be expressed via broadcast_in_dim.
2. Our compatibility guarantees haven't yet been finalized (see StableHLO Compatibility RFC v2 #115), so it's unclear what exactly this move will entail. E.g. are we going to keep both chlo.broadcast and stablehlo.broadcast around? For how long? How do we migrate users?
3. We don't yet have an RFC process to reason about changes to the StableHLO opset. Introduce StableHLO evolution process #196 was only introduced a few days ago and will still need some additional work.

As a result, we'll be focusing on completing the StableHLO v1.0 spec in the next few months, and by then we expect that all the blockers mentioned above will be resolved. Until then, I'll be unassigning this ticket.

[burmako]

Refactored this ticket to focus on ops which don't have equivalents in the HLO opset (not counting dynamism ops which are handled separately in #8). BatchNorm ops which were previously part of this ticket are now discussed in #603.