Single-state Aggregator and test refactor

[jmacd]

Simplify export.Aggregator usage

The Aggregator.Checkpoint API is renamed SynchronizedCopy and adds an argument, a different Aggregator into which the copy is stored. This replaces part of #799.

The SDK now stores two aggregators per synchronous record, one per asynchronous record.

The simple Integrator stores one additional aggregator per stateful record.

The constructors of all the Aggregators are modified to return a variable-length slice of aggregators. The SDK will allocate one or two of these at a time. The AggregationSelector.AggregatorFor() API now accepts ...*export.Aggregator to initialize one or more Aggregators with a single allocation.

Test refactoring

The Aggregator tests allocate 1, 2, or 4 Aggregators. This pattern is assisted by new2() and new4() constructors in the individual Aggregator tests. External testing is assisted by Unslice2() in the sdk/metric/aggregator/test package.

Removes a number of test AggregatorFor methods in favor of a single test test.AggregationSelector() used everywhere, where the metric name suffix indicates which class of Aggregator to use.

Minor Aggregator cleanups

The export.Aggregator contract is that Update() and SynchronizedCopy() are synchronized with each other. All the aggregation interfaces (Sum, LastValue, ...) are not meant to be synchronized, as the caller is expected to synchronize aggregators at a higher level after the Accumulator. Some of the Aggregators used unnecessary locking and have been cleaned up.

Several uses of metric.Number were replaced by int64 now that we use sync.Mutex in the MinMaxSumCount and Histogram Aggregators.

metric.Number(0) expressions are replaced by 0 where possible.

Removed alignment tests that were unnecessary because we are using sync.Mutex.

[MrAlias]

Looks good 👍

Nothing blocking, mostly just error checking and docstring updates.