Update get_mut() documentation to make clear how 'inner` can be used

[Byron]

This should help steering towards reset() in places where flush() + swap(get_mut, x) might otherwise be used.

It also mentions how to use get_mut() for streaming.

Fixes #555

[jongiddy]

This perhaps over-emphasizes the need for flushing. You can replace a Write implementation at any time without affecting the state of the Encoder/Decoder. To obtain the full output you must (at least logically) concatenate the output of all writers in order, possibly requiring a flush on each of them. However, it is not required to flush the Encoder/Decoder before each replacement.

Maybe something more like:

Suggested change

	/// The underlying writer may be mutated or replaced as long as this
	/// preserves the bytes and ordering of the logical output stream.
	///
	/// For streaming output, call [`flush`](Write::flush) before replacing the
	/// writer, such as with [`std::mem::take`], to retrieve all output produced
	/// so far. Concatenate output from each writer to reconstruct the complete
	/// stream.
	/// The underlying writer may be mutated or replaced as long as this
	/// preserves the bytes and ordering of the logical output stream.
	/// Concatenate output from each writer to reconstruct the complete
	/// stream.
	///
	/// To ensure output is retrieved as soon as it is available, call [`flush`](Write::flush)
	/// before replacing the writer.

[fintelia]

The part that neither of these descriptions really express is that calling flush changes the output bitstream by inserting a sync-flush. That's necessary to ensure that all input bytes passed in prior to the flush can be decoded without needing further bytes of output produced by any calls following the flush.

[Byron]

Thanks @jongiddy and @fintelia! Particularly the "flush has side-effects" note is something the caller should definitely be aware of.

What about this?

Encoders

    /// Acquires a mutable reference to the underlying writer.
    ///
    /// The underlying writer may be mutated or replaced as long as this
    /// preserves the bytes and ordering of the logical output stream.
    /// Concatenate output from each writer to reconstruct the complete stream.
    ///
    /// Replacing the writer does not require [`flush`](Write::flush). Call it
    /// first when all input accepted so far must be decodable without output
    /// from later writes. This inserts a sync-flush point and changes the output
    /// bitstream. This is useful before applying [`std::mem::take`] to
    /// [`get_mut`](Self::get_mut) when forwarding the stream incrementally.
    ///
    /// To start a new stream, use [`reset`](Self::reset); replacing the writer
    /// does not reset this encoder.

Decoders

    /// Acquires a mutable reference to the underlying writer.
    ///
    /// The underlying writer may be mutated or replaced as long as this
    /// preserves the bytes and ordering of the logical output stream.
    /// Concatenate output from each writer to reconstruct the complete stream.
    ///
    /// Replacing the writer does not require [`flush`](Write::flush). Call it
    /// first to write all decompressed output currently available to the
    /// current writer. This is useful before applying [`std::mem::take`] to
    /// [`get_mut`](Self::get_mut) when forwarding output incrementally.
    ///
    /// To start a new stream, call [`finish`](Self::finish) and create a new
    /// decoder; replacing the writer does not reset it.

I have also updated the PR to reflect the change everywhere, with both forms.