This works for WS2812B but not WS2812

[MartyMacGyver]

This doesn't work well on old 6-pin WS2812 devices, Strictly speaking, you can get it mostly working by using a WS2812_CYCLE time of about 240ns, but it's still glitchy.

Even then, a look with the logic analyzer reveals the problem: a bit-banged test using the Particle Photon and the Adafruit_NeoPixel library gives the following timings:

Low:   tH = 0.31µs ±0.01µs / tL = 0.94µs ±0.01µs
High:  tH = 0.82µs ±0.01µs / tL = 0.44µs ±0.01µs
Cycle: 1.25µs ±0.02µs

The ESP32 RMT timings (with WS2812_CYCLE = 240ns):

Low:   tH = 0.48µs ±0.01µs / tL = 0.47µs ±0.01µs
High:  tH = 0.95µs ±0.01µs / tL = 0.95µs ±0.01µs
Cycle: 1.43µs ±0.48µs

The spec timings are as follows:

WS2812:

Low:   tH = 0.35µs ±0.15µs / tL = 0.80µs ±0.15µs
High:  tH = 0.70µs ±0.15µs / tL = 0.60µs ±0.15µs
Cycle: 1.25µs ±0.60µs
Reset: >50µs

WS2812B:

Low:   tH = 0.35µs ±0.15µs / tL = 0.90µs ±0.15µs
High:  tH = 0.90µs ±0.15µs / tL = 0.35µs ±0.15µs
Cycle: 1.25µs ±0.15µs
Reset: >50µs

As a rule of thumb we see that a low bit needs a longer time in the low state than the high state, and vice versa for high. A 2:1 duty cycle should suffice for both, but a 50:50 duty cycle will not do (it's a wonder it works at all for the WS2812B let alone the WS2812 - presumably the tH is what's mediating this and so lengthening that timing serves to be decoded as low versus high.

If possible, the RMT should be driving a more appropriate signal. Otherwise it's not a very reliable solution versus bit-banging (which works remarkably poorly on the ESP32).

(Note: there's a lot of leeway when signalling, but not enough to allow 50:50 timing to work glitch-free.)