feature - implement RFC 023 approximate aggregates (#40)

[dannymeijer]

Summary

Implements the portable RFC 023 approximate aggregate slice from current main:

• Adds approx_count_distinct(...) and approx_percentile(..., accuracy=10000) as registry-backed aggregate helpers.
• Adds approximation policy metadata and stable InQL Substrait extension anchors.
• Carries aggregate argument lists through Prism/Substrait so approx_percentile keeps percentile and accuracy parameters.
• Keeps DataFusion implementation-name rewrites adapter-local (approx_distinct, approx_percentile_cont) without changing emitted InQL Substrait names.
• Adds Draft RFC 025 as the typed sketch logical value follow-up, linked to RFC 025: Typed sketch logical values #51, so RFC 023 does not smuggle sketch state in as strings/bytes.

Replaces closed stale PR #50 with a clean branch based on the merged RFC 022 mainline.

Type of change

• Bug fix
• New feature
• Refactor / maintenance
• Documentation
• CI / tooling
• RFC (adds/updates docs/rfcs/*)

Area(s)

• Package & tests
• Specification (RFCs)
• Documentation
• Automation & repo config
• Other

Key details

• User-facing behavior: users can build approximate grouped aggregates with approx_count_distinct(expr) and approx_percentile(expr, percentile, accuracy=10000). Approximate percentile output names include percentile and accuracy so distinct estimates over the same input column do not collide.
• Internals: aggregate measures now carry explicit argument lists; Prism structural equality compares those arguments; Substrait aggregate lowering uses the argument list rather than rebuilding from the primary expression; DataFusion consumer-plan adaptation only rewrites extension declarations at the adapter boundary.
• Risks: DataFusion approximate aggregate behavior is implementation-dependent, so docs describe approximation semantics without promising bit-for-bit portability. Sketch-state helpers remain RFC 025 Draft rather than lowerable RFC 023 functions.

Testing / verification

• make fmt INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• incan test tests/test_approximate_functions.incn
• incan test tests/test_function_registry.incn
• incan test tests/test_substrait_plan.incn
• incan test tests/test_prism.incn
• incan test tests/test_session_aggregates.incn
• make fmt-check INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• make test-style
• make registry-metadata INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• make build INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• make test INCAN=/Users/danny/Development/encero/incan/target/debug/incan (186 passed)
• make smoke-consumer INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• make pre-commit INCAN=/Users/danny/Development/encero/incan/target/debug/incan
• Manual verification described below

Manual verification notes:

• Rebuilt local Incan from merged #718 before testing; smoke-consumer needed a network-enabled rerun after sandbox DNS blocked crates.io lockfile generation.
• Review reports were recorded in the local InQL worktree under .agents/state/ and not in the central Incan findings corpus.

Docs impact

• No docs changes needed
• Docs updated
• Docs follow Divio intent (tutorial/how-to/reference/explanation) where applicable

If docs updated:

• Link(s): docs/language/reference/functions/approximate.md, docs/language/reference/builders/aggregates.md, docs/rfcs/023_approximate_sketch_functions.md, docs/rfcs/025_typed_sketch_logical_values.md, docs/rfcs/README.md, docs/release_notes/v0_1.md

Checklist

• I kept public docs user-focused and moved internals to contributing docs when appropriate
• I avoided duplicating canonical install/run instructions in multiple places
• I added/updated tests where it materially reduces regressions

Closes #40.
Refs #51.