Optimize MessageBus::Implementation#decode_channel_name method

[moberegger]

Another hotspot showing up in our production profiles.

This PR replaces the channel.split(ENCODE_SITE_TOKEN) with a two-branch approach. It first uses byteindex (to skip character-encoding boundary checks) to check for the "$|$" site token.

When no site token is found:

• Skip the split entirely and simply return [channel, nil]. This eliminates both the array and substring allocations that split would produce.

When a site token is found:

• Use the byteindex token position, use byteslice to extract the channel and site_id directly. This is faster than split because it scans the string only once and I believe also skips character-encoding boundary checks. byteslice preserves the source string's encoding, so non-ASCII channel names should be handled correctly.

Some benchmarks against decode_channel_name.

Without a site ID

ruby 4.0.2 (2026-03-17 revision d3da9fec82) +YJIT +PRISM [arm64-darwin25]

Channel: "/global/test"

=== Iterations per second (global channel, no site_id) ===
Warming up --------------------------------------
            original   871.044k i/100ms
           optimized     1.407M i/100ms
Calculating -------------------------------------
            original     10.053M (± 2.2%) i/s   (99.48 ns/i) -     50.521M in   5.028068s
           optimized     18.492M (± 3.2%) i/s   (54.08 ns/i) -     92.844M in   5.026627s

Comparison:
            original: 10052561.9 i/s
           optimized: 18491743.0 i/s - 1.84x  faster

=== Memory allocation (global channel, no site_id) ===
Calculating -------------------------------------
            original    80.000  memsize (     0.000  retained)
                         2.000  objects (     0.000  retained)
                         1.000  strings (     0.000  retained)
           optimized    40.000  memsize (     0.000  retained)
                         1.000  objects (     0.000  retained)
                         0.000  strings (     0.000  retained)

Comparison:
           optimized:         40 allocated
            original:         80 allocated - 2.00x more

With a site ID

ruby 4.0.2 (2026-03-17 revision d3da9fec82) +YJIT +PRISM [arm64-darwin25]

Channel: "/test/channel$|$default"

=== Iterations per second (channel with site_id) ===
Warming up --------------------------------------
            original   569.231k i/100ms
           optimized   677.742k i/100ms
Calculating -------------------------------------
            original      6.503M (± 2.4%) i/s  (153.77 ns/i) -     33.015M in   5.079579s
           optimized      7.503M (± 2.4%) i/s  (133.27 ns/i) -     37.954M in   5.061283s

Comparison:
            original:  6503391.7 i/s
           optimized:  7503464.3 i/s - 1.15x  faster

=== Memory allocation (channel with site_id) ===
Calculating -------------------------------------
            original   120.000  memsize (     0.000  retained)
                         3.000  objects (     0.000  retained)
                         2.000  strings (     0.000  retained)
           optimized   120.000  memsize (     0.000  retained)
                         3.000  objects (     0.000  retained)
                         2.000  strings (     0.000  retained)

Comparison:
            original:        120 allocated
           optimized:        120 allocated - same