refactor: remove predicate fusion and pushdown

[cpcloud]

This PR removes predicate pushdown and fusion in ibis/expr/analysis.py
in an effort to greatly simplify expression construction and increase
maintainability.

This is the first in a series of backwards-incompatible changes that
will culminate in the simplication of the core relational operators (#2919).

There are a number of benefits to this:

1. Fewer lines of code to maintain
2. Fewer lines of highly complex code to maintain
3. Much easier to debug broken queries
4. New contributors can more easily understand the library

We can revisit adding fusion back later, after designing a more robust and
modular API for optimizing at expression construction time.

There's one major breaking change:

1. Filter expressions that refer to a non-immediate child table are no longer
   valid and will raise an error (projections are almost all still valid,
   removing projection fusion will come in a follow-up).

   An example of this is:

t = ibis.table([("a", "string"), ("b", "int64")])
t[t.a == "foo"][t.b == 1]

previously this would fuse t.a == "foo" and t.b == 1 into a conjunction
and return t[(t.a == "foo") & (t.b == 1)].

The behavior is still available through other means, like the & operator, passing
a list of predicates, and if you must chain, you need to use a lambda.

Non-breaking changes:

1. SQL queries contain more nesting. Queries that compose a lot of filters will
   contain more subqueries.

   I don't view this as problematic, as most database engines having fairly
   comprehensive subquery unnesting capabilities.

[github-actions]

Unit Test Results

       19 files         19 suites   1h 41m 29s ⏱️
10 774 tests   8 717 ✔️   2 044 💤 13 ❌
54 565 runs  43 713 ✔️ 10 778 💤 74 ❌

For more details on these failures, see this check.

Results for commit 379d005.

♻️ This comment has been updated with latest results.

[icexelloss]

I am nervous about this. We use pandas backends pretty heavily - are there any perf regressions if we remove fusion?

[cpcloud]

This is still a WIP (hence the draft).

There will probably be more temporaries created.

I may ultimately close this PR in favor of a different approach that would allow us to use rewrite rules to keep most optimizations.

[cpcloud]

Closing this out for now.