Arrays as a Logical Plan

[gatesn]

This has been written in various places before, but I want to consolidate the plan with the status of progress and tracking issues.

This change is about making Vortex arrays represent the decompression logic required to get back uncompressed data, rather than now where they represent a mix of logic and data and much of the API has unpredictable performance characteristics.

The gist of the change is:

• Arrays represent a logical decompression plan.
• Expressions can be applied to arrays in constant time (by wrapping up an existing Array in a new ScalarFnArray)
• Array::optimize applies optimization rules to the array tree that do not require decompression any data (e.g. pushing a scalar function into the values of a run-end array)
• Array::execute represents CPU execution of an array into full decompressed form. Currently this is a Vortex Vector, although we also have first-class support for Array::execute_arrow to return an arrow-rs Array.

• Deprecate ComputeFn trait in favor of ScalarFn + AggregateFn
  • ScalarFn
  • ScalarFnArray
  • Remove existing scalar ComputeFns
  • AggregateFn
  • AggregateFnArray
  • Remove existing aggregate ComputeFns
• Optimization Framework
  • Array::optimize
  • Array::optimize_parent
• Execution Framework
  • Array::execute
  • Array::execute_parent
• Deprecate Array::evaluate in favor of Array::execute
• Add FilterArray to represent a lazy filter.
  • Allow filter mask to be either Mask or Array.
• Add SliceArray to represent a lazy slice.
• Add functions to Array vtable for common push-down operations
  • Slice, Filter, Take, Validity (is_not_null), to return Option<ArrayRef> without execution, or None if not possible.
• All arrays to return Validity without execution.
  • Do we need ScalarFnValidityArray to execute only the validity portion of a scalar function?
• Deprecate direct access to data on arrays.
  • Arrays to hold BufferHandles instead of Buffers.
  • CPU ExecutionCtx resolves BufferHandle to Buffer allowing access to data only during execution.
  • Remove access to buffers from Primitive, Boolean and other arrays.
• Deprecate non-execute functions on Array trait that require data access
  • is_valid, is_invalid, scalar_at, valid_count, invalid_count
  • append_to_builder
• Deprecate vortex-scalar in favor of typed wrappers around the more efficient vortex-vector/Scalar
• Add execution APIs
  • execute_vector, execute_datum
  • execute_arrow
  • execute_canonical
  • execute_scalar (?)
• All execution functions to require either a Session or an ExecutionCtx. ExecutionCtx::buffer(BufferHandle) -> Buffer. This prevents accidental execution outside of an execution context.
• Matcher API for downcasting arrays