Allow injecting reward and controller classes into AEES

[davidkfoss]

Problem

AEES.__init__ hardcodes the reward and controller construction. Even though BaseReward and BaseController are public protocols, every user gets:

• NormalizedLossImprovementReward
• DiscountedUCBController / BucketedContextualController / RandomController

There is no injection point, so the protocol abstraction is half-finished — the types are public but the instantiation is closed.

Proposal

Accept optional pre-built controller / reward instances (or factories) on AEES.__init__, falling back to today's defaults when not provided. Something like:

AEES(
    optimizer,
    reward: BaseReward | None = None,
    lr_controller: BaseController | None = None,
    noise_controller: BaseController | list[BaseController] | None = None,
    ...
)

Acceptance criteria

• Users can pass custom BaseReward / BaseController implementations.
• Defaults remain unchanged when nothing is passed (no behavior regression).
• Validation still rejects mismatched controller / context-mode combinations.
• Docs/example updated to show how to plug in a custom controller.