Make L2Denorm an array encoding instead of a scalar function

[connortsui20]

Rationale for this change

(semi-related) Tracking Issue: #9129

I would like to propose a rule we hold going forward: a scalar function never verifies its inputs, an encoding may. A scalar function takes anything well-typed and checks nothing past dtypes, which is all return_dtype can see anyway. An encoding owns a physical layout, so it gets VTable::validate and its own constructors, and it can hold invariants between its children. The corollary is the useful half: if you want to scan the data at construction time, you want an encoding.

Note that the rule only runs one way. FoR is the edge case I can think of, since encoded + reference is an ordinary scalar function over any integer column and its validate is purely structural. Nothing checks that the encoded values really are offsets from the minimum, because that is a compression heuristic and not a correctness requirement. So needing verification forces an encoding, but not needing it does not force a scalar function.

L2Denorm broke the rule in the direction that matters: its constructor took an ExecutionCtx and scanned both children for the unit-norm invariant, which ScalarFnFactoryExt has nowhere to put. Everything else follows from that. L2Norm reads its stored norms instead of recomputing, CosineSimilarity and InnerProduct reach into its physical children, L2DenormScheme already listed it in produced_encodings, and persisting it needed a ScalarFnArrayPlugin behind an env var. The operation itself (normalized * norm) is valid for any tensor and any float column, so the invariant only ever existed to justify encoding-specific optimizations.

What changes are included in this PR?

L2Denorm becomes a VTable in vortex-tensor/src/encodings/l2_denorm/ with two slots (normalized, norms).

• try_new scans for the exact invariants, try_new_trusted skips the scan for lossy normalized children whose stored norms stay authoritative. Neither is unsafe, since breaking the contract gives wrong answers rather than undefined behavior.
• Slice and filter push down into both children. Filter is new, since the generic ScalarFnArray rule only fires when at most one child is non-constant.
• L2DenormScheme cascades its children like TemporalScheme, so both HACK TO SUPPORT L2 DENORMALIZATION cases come out of CascadingCompressor. They existed because a scheme producing a ScalarFnArray had no compressible children to be sized against.

What APIs are changed? Are there any user-facing changes?

Before	After
scalar_fns::l2_denorm::*	encodings::l2_denorm::*
L2Denorm::try_new_array(n, s, ctx)	L2Denorm::try_new(n, s, ctx)
unsafe L2Denorm::new_array_unchecked(n, s)	L2Denorm::try_new_trusted(n, s)
L2Denorm::new()	removed, there is no scalar function
session.scalar_fns().register(L2Denorm)	session.arrays().register(L2Denorm), unconditional
ScalarFnArrayPlugin::new(L2Denorm)	removed

VX_SCALAR_FN_ARRAY_TENSOR_PLUGIN no longer gates L2Denorm, since the compressor can emit it and readers have to be able to open those files. It still gates the other three tensor plugins. The array ID and the L2DenormMetadata message are unchanged, so the on-disk bytes match what the plugin wrote. L2DenormScheme stays out of ALL_SCHEMES, because wiring it in changes default compression for every tensor column and wants its own numbers.