Filter out non-APDU interfaces in UsbTransport and other changes related to issue #12

[ImplOfAnImpl]

Edit: initially I wanted to create several PRs on top of each other, but it seems that GitHub can't do that when PRs come from a fork. So I've put all my changes into a single PR.

I have 4 commits here:

1. "Filter out non-APDU interfaces in UsbTransport".

   This addresses the "USB interface selection" part of A few issues related to device selection #12.

2. "Improve BLE device discovery, support new models".

   This addresses the "Bluetooth device selection" part of A few issues related to device selection #12. Also, Flex and NanoGen5 were added to the list of models.

   I've only checked BT connectivity on Nano Gen 5, but since the corresponding UUIDs come from Ledger Live source code, I'm pretty sure it should work for other models too.
   Though additional testing won't hurt of course.

3. "Replace Error::Unknown with more specific errors".

   Not related to A few issues related to device selection #12. It's just that Error::Unknown is way too vague, so I've added a few more specific variants instead. Also, a couple of Error variants were unused, so I removed them.

4. "UsbTransport and UsbDevice are no longer Send. Remove explicit "impl !Send" and "impl Send" for UsbTransport/UsbDevice".

   This addresses the last item in A few issues related to device selection #12 - UsbTransport/UsbDevice not being Send but being marked as Send anyway.

   As I've said in the issue, IMO UsbTransport/UsbDevice should be Send. But since I haven't got a reply to that, I decided to go the other way and actually make them non-Send.
   So, ledger-lib no longer uses the async-trait crate; instead, when the Send bound on returned futures is not needed, async functions are declared directly and the corresponding warning
   ("use of async fn in public traits is discouraged as auto trait bounds cannot be specified") is silenced via #[allow(async_fn_in_trait)].
   If the Send bound is needed, then the more verbose syntax fn foo() -> impl Future<...> + Send is used.

   Note that there is also the trait-variant crate whose make macro allows to have both the normal "async fn" syntax in traits and the Send bound on the futures.
   But it looks like the crate is unmaintained and make was producing some weird errors in cases when a trait method had a body. So I guess the verbose syntax is the way.

   I also had to introduce an additional trait NonSendExchange, which is like Exchange but without the Send bound. It's needed because UsbDevice can't implement Exchange.

   Also, just in case, I bumped minimum hidapi version to "2.6" (the previously referenced v2.1 required that only one instance of HidApi can exist at time, this restriction was lifted in later versions).

   P.S. async-trait is still used in other places. Probably I should remove it there too, for consistency?

[ImplOfAnImpl]

FYI: changes in this file are only needed to be able to override the backend from command line when building ledger-cli (e.g. cargo run --no-default-features --features transport_usb_hidraw --bin ledger-cli -- list)

[ImplOfAnImpl]

FYI: I had a situation on my Ubuntu where HidApi::devices_list only returned the FIDO interface and not the APDU one when using the hidraw backend, but it returned both interfaces when using libusb. The issue was solved by physically disconnecting and reconnecting the device. So it looks like hidraw is unreliable in general and not only with WSL.

[yogh333]

LGTM, could you please run clippy in order to make tCI happy? Thanks

[ImplOfAnImpl]

could you please run clippy in order to make tCI happy?

Done

[ImplOfAnImpl]

Are there any plans to merge this?

[ImplOfAnImpl]

I've pushed one more commit, making futures returned by Device's methods Send, otherwise it becomes unusable in the "normal" async code. This, on the other hand, makes Device unusable with types that only implement NonSendExchange (i.e. UsbDevice), which is kind of ugly. I guess just making UsbTransport/UsbDevice Send would have been a better way.

Also, ledger-lib now uses the async-trait macro again, to work around a compiler bug.

Also, I've updated min Rust version to 1.85, because it's what one of dependencies needs.