Move MentatError away from the top-level crate

[grigoryk]

This PR is in preparation for landing some of the work in #563 - specifically, the Pre work which splits transaction-like functionality away from the top-level into its own inner crate.

Pre commits move any trait or type that is necessary to compose an error out of a given crate, and into a "leaf" crate, which use the *-traits naming convention (e.g. db and db-traits). This pattern is similar to how Servo structures their internal inter-dependencies, although at a weaker level (since I'm retrofitting this pattern onto the current world, and going all-in on it would be quite a bit of work!).

This move is needed so that we can use a single Error type in our public API.

Without these changes, the Pre patch above would have to leak a TransactionError into mentat's public API space, which is a subset of the MentatError. If the MentatError is defined in an inner crate, it can now be referenced in other inner crates, in absence of circular dependencies.

This pattern is useful for crates which:

• tie a lot of internal functionality together
• are exposed as part of the public API
• themselves don't introduce new error types

A prime example is the "mentat transaction state manager" (the transaction crate in #563), and the upcoming tolstoy which makes heavy use of the transaction crate.

[grigoryk]

This introduces a temporary inconsistency in how crates are named - the new crates I've added don't use the "mentat_" prefix. The intention is to remove that prefix entirely, but that will be done once work that depends on this - in #563 - lands, to avoid an even more annoying rebase.

[grigoryk]

I don't expect the public API to have changed - but it's probably a fair statement that it did in some subtle way, and we currently don't have much in place to ensure that it didn't change. Regarding the lack of src/ in the subcrates, it just seemed a bit neater to me.