Commit aa43ed8 was fixing the light sleep overhead calculation for
ESP32-S2. However it also changed the overhead values for ESP32,
resulting in incorrect light sleep time. This caused regression in
light sleep example test.
Revert the original values for the ESP32, keep different set of values
for each chip.
Based on the original code (`esp32/sleep_modes.c`), `RTC_SLEEP_PD_XTAL`
is always given as an argument to `esp_sleep_start`. Enforce this in
function to get power down flags to avoid redundancy.
