…o the formatter as kwargs

… context

Co-authored-by: nicholas a. evans <nick@rubinick.dev>

The biggest change is simplifying to _only_ support hash context.
Subclasses may use other context types, or imbue specific hash keys with
special behavior.

Logger:
* Only support hash context in `#merge_context` and `#format_message`.
* Add public `#context` method.
* Add private `#context_store` method (for misbehaving subclasses).
* Add rdoc for new and updated public methods
* Add `context` keyword parameter to other level-specific methods
* Simplify `#format_message` by delegating to `#merge_context`.
* Handle nil `prev_context` in `#merge_context`
* Move the `context.dup` from `#merge_context` into `#with_context`,
  to avoid `context.dup` in `#format_message`.  Context is also frozen,
  so `#context` can be safe without dup.

Formatter:
* Only support hash context in `#format_context`.
* Print context after the message.
* Context values are escaped.
* Don't use [brackets] around context.
* Fix backward compatibility with ruby < 2.7 (`#filter_map`)
* Remove trailing whitespace from `#format_context`.
* Adds several more methods as extension points: `#format_pair`,
  `#format_value`, `#join_fields`, and `#format_message_with_context`.

Tests:
* Update for the changes listed above.
* Fix backward compatibility with ruby < 2.7 (kwargs handling)

I want each logger instance to have its own default context in addition
to the fiber-scoped overrides.  With that in mind, I think it's better
to name the hash of overrides differently.