1. Top-down trait request
2. Bottom-up trait derivation
3. Trait enforcement without AbstractConverter

How to use?

1. Enable top-down optimization by setting {VolcanoPlanner#setTopDownOpt(boolean)}
or add 'calcite.planner.topdown.opt=true' to saffron.properties config file.

2. Let your convention's rel interface extends {PhysicalNode}, see
{EnumerableRel} as an example.

3. Each physical operator overrides any one of the two methods:
{PhysicalNode#passThrough(RelTraitSet)} or
{PhysicalNode#passThroughTraits(RelTraitSet)} depending on your needs.

4. Choose derive mode for each physical operator by overriding
{PhysicalNode#getDeriveMode()}.

5. If the derive mode is {DeriveMode#OMAKASE}, override method
{PhysicalNode#derive(List)} in the physical operator, otherwise, override
{PhysicalNode#derive(RelTraitSet, int)} or
{PhysicalNode#deriveTraits(RelTraitSet, int)}.

6. Mark your enforcer operator by overriding {RelNode#isEnforcer()}, see
{Sort#isEnforcer()} as an example. This is important, because it can help
{VolcanoPlanner} avoid unnecessary trait propagation and derivation, therefore
improve optimization efficiency.

7. Implement {Convention#enforce(RelNode, RelTraitSet)} in your convention,
which generates appropriate physical enforcer. See
{EnumerableConvention#enforce(RelNode, RelTraitSet)} as example. Simply return
null if you don't want physical trait enforcement.

How does it work?

Let S# denote the seed physical operator in a RelSet after logical and physical
rules transformation, P# denote the physical operator generated by passing down
parent trait requirements, D# denote the physical operator generated by
deriving from child delivered traitSets.

The initial rel list state in a RelSet is as follows:
    cursor
      |
      V
     S1, S2

When we create a task for RelSubset1, the task will immediately pass the
subset's traitSet to seed operators, S1 and S2, now we have:
    cursor
      |
      V
     S1, S2, P1, P2

The subset task will create a optimization task for the relnode pointed by
cursor, and move cursor to next available physical operator S2. In the task for
S1, it will continue optimize its child nodes, which are RelSubsets. After
child inputs optimization is finished, S1 will derive new relnodes from
delivered subsets in input RelSet. Once task for S1 is completed, we have:
        cursor
          |
          V
     S1, S2, P1, P2, D1

The subset task continues scheduling task for S2, P1... until there is no more
relnode created for the RelSet, then we have:
                                cursor
                                  |
                                  V
     S1, S2, P1, P2, D1, D2, D3, null

When a task for another RelSubset2 is created, the task will try to pass down
the subset's traitSet to seed operator S1 and S2, now the RelSet looks like:
                                cursor
                                  |
                                  V
     S1, S2, P1, P2, D1, D2, D3, P3, P4

The process continues till there is no more subsets or relnodes created for the
RelSet.

See https://lists.apache.org/thread.html/r38ea71968c069f465921e7197488329c15413b46831c90ad4d48f87e%40%3Cdev.calcite.apache.org%3E

Close apache#1953

…Gen Implementation[CALCITE-3224]"

This reverts commit 89cb40e.