Fix SH1106 issue when there are other i2c devices on the line and tra…

[zefir-o]

…nsmition fails and it causes issues on the screen

[marcelstoer]

The core idea is sound: the old code wrote to buffer_back before confirming the display received the data,
so a failed transmission silently suppressed retries on the next display() call. Deferring the sync until after confirmed success is the right fix.

Dead #ifdef nested inside its own #else branch (non-double-buffer path)

#else   // <-- OLEDDISPLAY_DOUBLE_BUFFER is NOT defined here                                       
  ...
  #ifdef OLEDDISPLAY_DOUBLE_BUFFER   // always false — dead code
    if (txOk) memcpy(buffer_back, buffer, displayBufferSize);
  #endif                                                                                                                                                                              
#endif

From what I understand, OLEDDISPLAY_DOUBLE_BUFFER can never be defined inside its own #else block, so the memcpy is unreachable. txOk tracking in the non-double-buffer path is therefore a complete no-op.
This block should be removed (and since buffer_back doesn't exist in that path, removing should be ok).

Other issues

1. sendDataChunk writes byte-by-byte instead of using _wire->write(const uint8_t*, size_t). Most Wire
   implementations support bulk writes, which are more efficient:

// current                                                                                                                                                                            
for (uint8_t i = 0; i < length; i++) _wire->write(data[i]);
// suggested                                                                                                                                                                          
_wire->write(data, length);

2. No retry in sendDataChunk: sendCommandChecked retries twice on failure, but sendDataChunk does not.
   Maybe worth a comment?
3. pageOk short-circuit ordering is correct (both sendCommandChecked calls in the address phase are
   always executed because each is the left operand of &&). A brief comment would help.

Kudos

• Replacing the hardcoded 8 and 128 in the non-double-buffer loop with displayHeight / 8 and displayWidth is
  a correctness improvement for non-128×64 geometries.
• The updated comment ("Do not update buffer_back yet: if an I2C transfer fails, the same dirty region must be retried.") clearly captures the invariant.
• yield() is preserved in all exit paths through the page loop — both the continue on command failure and the break on data chunk failure flow through the outer yield().

[zefir-o]

@marcelstoer thank you for reviewing the PR.
I updated it, take a look, please.

[marcelstoer]

LGTM, thanks

[zefir-o]

@marcelstoer thank you for reviewing the PR.
I don't see the merge button. Should it be performed from your side?