feat(expander): snapshot INTF / INTCAP at begin() for deep-sleep wake source detection

[lanrat]

IOExpander::begin() does a 22-byte sequential bulk read of MCP23017
registers 0x00..0x15. That sequence passes through INTFA/B (0x0E,
0x0F) before INTCAPA/B (0x10, 0x11) — and on the MCP23017, reading
INTCAP clears the INTF latch as a side effect.

So after begin() returns, the live INTF register reads as 0, even
when a wake-triggering interrupt was actually pending. Applications
using the MCP's INT line to wake the host from deep sleep can't
determine which pin caused the wake. v11.0.1's getIntState() reads
INTCAP fresh, but INTCAP can be overwritten by any new interrupt event
between begin() and the call (e.g. capacitive noise during boot).
The only reliable moment to capture the wake source is during begin()
itself.

Fix

Snapshot the cached pre-clear INTF/INTCAP values into dedicated members
in begin(), expose via two new public methods:

uint16_t getInterruptFlagsAtBegin();    // INTFA in low, INTFB in high
uint16_t getInterruptCaptureAtBegin();  // INTCAPA in low, INTCAPB in high

Callers pair them to distinguish rising vs falling edges:

uint16_t intf    = display.expander1.getInterruptFlagsAtBegin();
uint16_t intcap  = display.expander1.getInterruptCaptureAtBegin();
uint16_t rising  = intf &  intcap;   // press
uint16_t falling = intf & ~intcap;   // release

Verification

Inkplate 10 V1 with setIntPin(_, RISING) on the three touchpads.
Pressing each pad produces the expected INTFB/INTCAPB bit set;
release-edge wakes show INTFB != 0 but INTCAPB == 0 and are
correctly filtered out by the application.

Compat / scope

Strictly additive — no existing API or behavior is modified. Two new
public methods, 4 bytes of state per instance. Same file gating as the
existing MCP code (ARDUINO_INKPLATE6 || ARDUINO_INKPLATE10 || ARDUINO_INKPLATE6PLUS).