Fix external RTC detection on Port A and the sleep wakeup handling

[ainyan03]

Three independent fixes found while going through the open issues. Each one is a separate commit.

1. External RTC on Port A is never detected (#266)

RTC_Class::begin() passes the given I2C_Class* to the RX8130 instances, but not to the PCF8563 fallback, whose third parameter defaults to &In_I2C. With cfg.external_rtc = true the call M5.Rtc.begin(&M5.Ex_I2C) therefore probes the internal bus, and UNIT-RTC on Port A is never found.

The argument was present in v0.2.10 and was dropped in e35a905 (first released in v0.2.11), which matches the reporter's version bisection - including why M5Grey is the one board that still works: its internal bus is G21/G22, the same pins as Port A.

2. deepSleep() aborts on M5PaperS3 when touch_wakeup is requested (#205)

On targets without EXT0 support, deepSleep() enabled EXT1 wakeup unconditionally inside ESP_ERROR_CHECK. The PaperS3 touch INT pin is GPIO48, which is not an RTC IO, so esp_sleep_enable_ext1_wakeup_io() returns ESP_ERR_INVALID_ARG and the call aborts instead of sleeping. lightSleep() already handled this case with gpio_wakeup_enable().

The pin is now validated with rtc_gpio_is_valid_gpio(), and a warning is logged when it cannot be used as a deep sleep wakeup source.

The same block also built the EXT1 pin mask from _wakeupPin while the wait loop used wpin, so the M5PaperMono override (GPIO4) never reached the mask; both use wpin now.

3. A sleep duration of zero silently means "sleep forever" (#285)

micro_seconds = 0 meant "no timer wakeup", and 0 was also the default argument - so deepSleep() with no arguments means "sleep until the wakeup pin triggers". A caller that computes a duration and accidentally ends up with 0 sleeps until an external reset, which is hard to tell apart from a freeze.

That intent now has its own value:

    static constexpr std::uint64_t sleep_no_timer = ~0ull;

    void deepSleep (std::uint64_t micro_seconds = sleep_no_timer, bool touch_wakeup = true);
    void lightSleep(std::uint64_t micro_seconds = sleep_no_timer, bool touch_wakeup = true);

• calls without arguments behave exactly as before;
• an explicit 0 no longer sleeps and logs a warning instead. For deepSleep() the check runs before M5.Display.sleep(), so the caller resumes with the display still on;
• the sentinel is never handed to esp_sleep_enable_timer_wakeup() (about 580,000 years, which would be rejected) - it keeps taking the "disable the timer wakeup source" path;
• lightSleep() now also tracks whether a wakeup pin was actually enabled, and warns when neither a timer nor a pin wakeup source is in effect.

This is a breaking change for code that passes a literal 0 in order to sleep indefinitely: such code needs M5.Power.sleep_no_timer instead. In public GitHub code the no-argument form outnumbers explicit 0 by roughly 97 to 11, and the direction of the change is the safe one - the device stays awake and logs why, rather than never waking up.

Verification

Builds cleanly with PlatformIO for esp32 (EXT0 path), esp32s3 (EXT1 path) and the PC build (M5UNIFIED_PC_BUILD).

Not verified on hardware yet - deepSleep() on PaperS3 and UNIT-RTC on Port A would be the two worth checking.