Add expression balancing for xilinxhls backend

[francescobrivio]

PR Description

This PR is addressing issue https://github.com/thesps/conifer/issues/66. Main changes:

• In commit 099599d0a20a5a0ba7233a7a9bbbf3186eeb32a0 I am implementing a tree reduction method similar to what is done in hlt4ml (see these lines)
• The method is used when summing the tree scores in backends/xilinxhls/firmware/BDT_unrolled.h
• For this to work I had to reverse the indexes of the score_t scores array, from:

  score_t scores[n_trees][fn_classes(n_classes)];

  to

  score_t scores[fn_classes(n_classes)][n_trees];

These changes bring a notable improvements in latency (and partially in resource consumption) when using the AP_SAT flag in the BDT ScorePrecision, while maintaining performances identical to the current version of conifer in terms of BDT score.

Additionally, in commit bffd3a6125e31f68cce29ffefeb1a00d32343b8f I fixed a typo in the "Development instructions" and added pandas to the dev_requirements.txt file.

PR Validation

The following tests are done with:

• xgboost v1.7.5
• Vitis_LHS v2023.2
• Binary classification BDT with 327 trees and max_depth = 4

Latency/Resources table:

Conifer version	ScorePrecision	vsynth LUT	vsynth FF	Latency
master	ap_fixed<11,4,AP_RND_CONV,AP_SAT>	51936	5032	133
This PR	ap_fixed<11,4,AP_RND_CONV,AP_SAT>	51163	4489	9
--------	--------	--------	--------	--------
master	ap_fixed<11,4,AP_RND_CONV>	43329	3542	4
This PR	ap_fixed<11,4,AP_RND_CONV>	42988	3800	4

Score plots:

[thesps]

Thanks for this nice contribution! I've checked on some other models and see consistent improvements as you have shown. The tests are passing too (well, as much as they are passing on master branch...).

I will therefore merge this!