Martingale and predictable parts of sumRewards

[RemyDegenne]

Adapted from the auto-formalization at https://github.com/RemyDegenne/alpha-rar (with extensive changes, in particular generalization to arbitrary environments).

• generalize sumRewards to an arbitrary feedback type (with the required algebraic structure)
• add a definition for the indicator of an action being selected at a time
• add a definition for the mean of the distribution of an action at a time
• add martingale and predictable parts of sumRewards, and show that they indeed have those properties