ESP/Speaker: Avoid repeated heap allocations in `TAS5805::play()`

## Problem

`TAS5805::play()` allocates several large buffers each time it is called:

- approximately 32 KB DMA/staging buffer
- up to approximately 16 KB file buffer
- approximately 16 KB stereo I2S buffer

These buffers are then freed before `play()` returns.

Because `audio_task()` repeatedly calls `play()`, playback repeatedly allocates and frees large heap blocks. On a device that also performs networking, HTTP, TLS, and other dynamic allocations, this creates unnecessary heap churn and may contribute to long-term heap fragmentation.

The issue is not simply the total amount of allocated memory; repeatedly requesting and freeing large blocks can reduce the size of the largest contiguous free region over time.

## Proposed change

Allocate the playback buffers once and reuse them across playback passes.

Suitable approaches include:

- allocating them during TAS5805/audio initialization and retaining them for the lifetime of the player, or
- lazily allocating them on the first playback pass and reusing them thereafter.

The I2S buffer must continue to satisfy any memory/DMA requirements imposed by the I2S driver.

Release the buffers only when playback resources are explicitly torn down, such as during an appropriate shutdown/reinitialization path.

Start by locating TAS5805::play() and the audio_task() call path, then trace how playback and audio resources are initialized and torn down. Confirm that the DMA/staging, file, and stereo I2S buffers are reused across playback passes, still meet I2S memory requirements, and are released during explicit resource teardown.

Written by the indexing model from the issue text.