Improve representation of IR and metadata

[jamii]

Current design choices

• Use Let to express sharing
• Store metadata (types, keys) in RelationExpr::Get rather than look up the corresponding Let
• Recalculate metadata for other RelationExpr on the fly
• Mutate RelationExprs during transforms
• Do CSE as a final transform

Problems with the current design:

• Stored metadata gets out of sync (eg we mutate the value of a Let in a way that adds/removes keys, but don't update the corresponding Get)
• If we wanted to avoid storing metadata, when calling Get{..}.typ() we'd need some way to find walk up the tree to the corresponding Let
• Recalculating metadata is expensive and we have to code defensively to avoid performance problems
• Caching metadata is difficult because RelationExpr is often mutated
• Some metadata is only derived in the sql planner and can't be reconstructed (eg the scale/precision of decimals inside a ScalarExpr). If we want to eg pull out part of a Filter into a Map we can't figure out the correct ColumnType.
• Recent transforms to improve subqueries require inlining all Lets, so we lose the benefit of expressing sharing explicitly but still bear all the costs.

Two proposed changes.

Scalar type inference:

• record precision/scale in every ScalarExpr that computes with decimals, so that we can write ScalarExpr::typ(input_typ: RelationType) -> ScalarType
• remove ColumnType from Map

Sharing:

• Get is only used for importing views
• rather than Let, we express sharing by pointing all instances at the same RelationExpr
• the Unbind and InlineLet transforms are no longer needed
• transforms use map rather than visit_mut
• we do some kind of clever interning so that two RelationExpr nodes with the same inputs should always point to the same allocation
• now that RelationExprs are immutable we can cheaply cache metadata by comparing pointer value

Concerns:

• is passing around the intern thing un-ergonomic?
• will transforms break sharing horribly?

[jamii]

Everyone was :+1:

@benesch is taking the first change, @jamii the second.