Allow to box messages in the channels implicitly

[vstakhov]

This PR allows to box all messages when sending them over channels. The behaviour is controlled by attribute boxed_messages of the orchestra macro, for example:

#[orchestra(signal=SigSigSig, event=Event, gen=AllMessages, error=OrchestraError, boxed_messages=true)]
pub struct Orchestra {
	#[subsystem(consumes: MsgA, sends: [MsgB])]
	sub_a: AwesomeSubSysA,

	#[subsystem(consumes: MsgB, sends: [MsgA])]
	sub_b: AwesomeSubSysB,
}

This change is intended to be transparent for the orchestrated subsystems.

[drahnr]

I think the approach would work, my question would be it might be less of a 🔨 to have it on a per message type, such that one could either annotate the message enum that a subsystem consumes with a proc-macro / individual fields / the subsystem;

#[orchestra::giftwrap(all)]
enum FooMessage {
}

#[orchestra::giftwrap]
enum BarMessage {

#[ribbon]
X(Fat),
}

(obviously the naming has to be adjusted)

or

#[orchestra(signal=SigSigSig, event=Event, gen=AllMessages, error=OrchestraError, boxed_messages=true)]
pub struct Orchestra {
	#[subsystem(consumes: boxed MsgA, sends: [MsgB])]
	sub_a: AwesomeSubSysA,

	#[subsystem(consumes: boxed MsgB, sends: [MsgA])]
	sub_b: AwesomeSubSysB,
}

such that the consumer defines the behavior of transparent boxing.

Mostly ideas, feel free to discard, the impl works.

[vstakhov]

Benchmarks from my laptop (osx):

no jemalloc:

send boxed message u8   time:   [320.24 ns 328.55 ns 336.36 ns]
send unboxed message u8 time:   [285.09 ns 291.20 ns 296.45 ns]
send boxed message 1029 bytes time:   [738.99 ns 763.95 ns 788.70 ns] 
send unboxed message 1029 bytes time:   [615.58 ns 621.74 ns 628.62 ns]

preloaded jemalloc:

send boxed message u8   time:   [289.96 ns 297.19 ns 303.40 ns]
send unboxed message u8 time:   [252.18 ns 257.18 ns 261.48 ns]
send boxed message 1029 bytes time:   [386.75 ns 401.52 ns 415.99 ns]
send unboxed message 1029 bytes time:   [613.14 ns 619.94 ns 627.32 ns]

So almost no difference without jemalloc (boxed version is slightly slower in all cases) and significant difference when jemalloc is loaded.

[vstakhov]

@drahnr I think we need to decide something about this PR as it is the base for others :) I'd say that implicit boxing is a good thing for the polkadot case (jemalloc and async-channel). Per-message boxing that you suggest is not very much different from the case where an orchestra user specifies explicitly a type of message equal to Box<T> with the only exception that this case is totally broken if multiple subsystems want to consume Box<T>. However, the former bug is also out of the scope for the suggested PR.

[drahnr]

One dependency question.

Do we want to have an API to send i.e. a pre-boxed version of some messages?

Other than that LGTM (although I am not the greatest fan of the added complexity to orchestra. I'd have preferred to make it around the message types being sent, so orchestra can stay oblivious to the message type )

[vstakhov]

Do we want to have an API to send i.e. a pre-boxed version of some messages?

Hum, what do you mean? By default, orchestra does not perform implicit boxing - it must be specified explicitly in orchestra macro. I like your idea about boxing for individual subsystems and that might be another improvement I think.